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/syscalls.h>
32 #include <linux/timex.h>
33 #include <linux/jiffies.h>
34 #include <linux/cpuset.h>
35 #include <linux/export.h>
36 #include <linux/notifier.h>
37 #include <linux/memcontrol.h>
38 #include <linux/mempolicy.h>
39 #include <linux/security.h>
40 #include <linux/ptrace.h>
41 #include <linux/freezer.h>
42 #include <linux/ftrace.h>
43 #include <linux/ratelimit.h>
44 #include <linux/kthread.h>
45 #include <linux/init.h>
46 #include <linux/mmu_notifier.h>
47 
48 #include <asm/tlb.h>
49 #include "internal.h"
50 #include "slab.h"
51 
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/oom.h>
54 
55 static int sysctl_panic_on_oom;
56 static int sysctl_oom_kill_allocating_task;
57 static int sysctl_oom_dump_tasks = 1;
58 
59 /*
60  * Serializes oom killer invocations (out_of_memory()) from all contexts to
61  * prevent from over eager oom killing (e.g. when the oom killer is invoked
62  * from different domains).
63  *
64  * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
65  * and mark_oom_victim
66  */
67 DEFINE_MUTEX(oom_lock);
68 /* Serializes oom_score_adj and oom_score_adj_min updates */
69 DEFINE_MUTEX(oom_adj_mutex);
70 
is_memcg_oom(struct oom_control * oc)71 static inline bool is_memcg_oom(struct oom_control *oc)
72 {
73 	return oc->memcg != NULL;
74 }
75 
76 #ifdef CONFIG_NUMA
77 /**
78  * oom_cpuset_eligible() - check task eligibility for kill
79  * @start: task struct of which task to consider
80  * @oc: pointer to struct oom_control
81  *
82  * Task eligibility is determined by whether or not a candidate task, @tsk,
83  * shares the same mempolicy nodes as current if it is bound by such a policy
84  * and whether or not it has the same set of allowed cpuset nodes.
85  *
86  * This function is assuming oom-killer context and 'current' has triggered
87  * the oom-killer.
88  */
oom_cpuset_eligible(struct task_struct * start,struct oom_control * oc)89 static bool oom_cpuset_eligible(struct task_struct *start,
90 				struct oom_control *oc)
91 {
92 	struct task_struct *tsk;
93 	bool ret = false;
94 	const nodemask_t *mask = oc->nodemask;
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_in_oom_domain(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
oom_cpuset_eligible(struct task_struct * tsk,struct oom_control * oc)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  * kthread_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  */
find_lock_task_mm(struct task_struct * p)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  */
is_sysrq_oom(struct oom_control * oc)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. */
oom_unkillable_task(struct task_struct * p)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  * Check whether unreclaimable slab amount is greater than
173  * all user memory(LRU pages).
174  * dump_unreclaimable_slab() could help in the case that
175  * oom due to too much unreclaimable slab used by kernel.
176 */
should_dump_unreclaim_slab(void)177 static bool should_dump_unreclaim_slab(void)
178 {
179 	unsigned long nr_lru;
180 
181 	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
182 		 global_node_page_state(NR_INACTIVE_ANON) +
183 		 global_node_page_state(NR_ACTIVE_FILE) +
184 		 global_node_page_state(NR_INACTIVE_FILE) +
185 		 global_node_page_state(NR_ISOLATED_ANON) +
186 		 global_node_page_state(NR_ISOLATED_FILE) +
187 		 global_node_page_state(NR_UNEVICTABLE);
188 
189 	return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
190 }
191 
192 /**
193  * oom_badness - heuristic function to determine which candidate task to kill
194  * @p: task struct of which task we should calculate
195  * @totalpages: total present RAM allowed for page allocation
196  *
197  * The heuristic for determining which task to kill is made to be as simple and
198  * predictable as possible.  The goal is to return the highest value for the
199  * task consuming the most memory to avoid subsequent oom failures.
200  */
oom_badness(struct task_struct * p,unsigned long totalpages)201 long oom_badness(struct task_struct *p, unsigned long totalpages)
202 {
203 	long points;
204 	long adj;
205 
206 	if (oom_unkillable_task(p))
207 		return LONG_MIN;
208 
209 	p = find_lock_task_mm(p);
210 	if (!p)
211 		return LONG_MIN;
212 
213 	/*
214 	 * Do not even consider tasks which are explicitly marked oom
215 	 * unkillable or have been already oom reaped or the are in
216 	 * the middle of vfork
217 	 */
218 	adj = (long)p->signal->oom_score_adj;
219 	if (adj == OOM_SCORE_ADJ_MIN ||
220 			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
221 			in_vfork(p)) {
222 		task_unlock(p);
223 		return LONG_MIN;
224 	}
225 
226 	/*
227 	 * The baseline for the badness score is the proportion of RAM that each
228 	 * task's rss, pagetable and swap space use.
229 	 */
230 	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
231 		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
232 	task_unlock(p);
233 
234 	/* Normalize to oom_score_adj units */
235 	adj *= totalpages / 1000;
236 	points += adj;
237 
238 	return points;
239 }
240 
241 static const char * const oom_constraint_text[] = {
242 	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
243 	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
244 	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
245 	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
246 };
247 
248 /*
249  * Determine the type of allocation constraint.
250  */
constrained_alloc(struct oom_control * oc)251 static enum oom_constraint constrained_alloc(struct oom_control *oc)
252 {
253 	struct zone *zone;
254 	struct zoneref *z;
255 	enum zone_type highest_zoneidx = gfp_zone(oc->gfp_mask);
256 	bool cpuset_limited = false;
257 	int nid;
258 
259 	if (is_memcg_oom(oc)) {
260 		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
261 		return CONSTRAINT_MEMCG;
262 	}
263 
264 	/* Default to all available memory */
265 	oc->totalpages = totalram_pages() + total_swap_pages;
266 
267 	if (!IS_ENABLED(CONFIG_NUMA))
268 		return CONSTRAINT_NONE;
269 
270 	if (!oc->zonelist)
271 		return CONSTRAINT_NONE;
272 	/*
273 	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
274 	 * to kill current.We have to random task kill in this case.
275 	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
276 	 */
277 	if (oc->gfp_mask & __GFP_THISNODE)
278 		return CONSTRAINT_NONE;
279 
280 	/*
281 	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
282 	 * the page allocator means a mempolicy is in effect.  Cpuset policy
283 	 * is enforced in get_page_from_freelist().
284 	 */
285 	if (oc->nodemask &&
286 	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
287 		oc->totalpages = total_swap_pages;
288 		for_each_node_mask(nid, *oc->nodemask)
289 			oc->totalpages += node_present_pages(nid);
290 		return CONSTRAINT_MEMORY_POLICY;
291 	}
292 
293 	/* Check this allocation failure is caused by cpuset's wall function */
294 	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
295 			highest_zoneidx, oc->nodemask)
296 		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
297 			cpuset_limited = true;
298 
299 	if (cpuset_limited) {
300 		oc->totalpages = total_swap_pages;
301 		for_each_node_mask(nid, cpuset_current_mems_allowed)
302 			oc->totalpages += node_present_pages(nid);
303 		return CONSTRAINT_CPUSET;
304 	}
305 	return CONSTRAINT_NONE;
306 }
307 
oom_evaluate_task(struct task_struct * task,void * arg)308 static int oom_evaluate_task(struct task_struct *task, void *arg)
309 {
310 	struct oom_control *oc = arg;
311 	long points;
312 
313 	if (oom_unkillable_task(task))
314 		goto next;
315 
316 	/* p may not have freeable memory in nodemask */
317 	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
318 		goto next;
319 
320 	/*
321 	 * This task already has access to memory reserves and is being killed.
322 	 * Don't allow any other task to have access to the reserves unless
323 	 * the task has MMF_OOM_SKIP because chances that it would release
324 	 * any memory is quite low.
325 	 */
326 	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
327 		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
328 			goto next;
329 		goto abort;
330 	}
331 
332 	/*
333 	 * If task is allocating a lot of memory and has been marked to be
334 	 * killed first if it triggers an oom, then select it.
335 	 */
336 	if (oom_task_origin(task)) {
337 		points = LONG_MAX;
338 		goto select;
339 	}
340 
341 	points = oom_badness(task, oc->totalpages);
342 	if (points == LONG_MIN || points < oc->chosen_points)
343 		goto next;
344 
345 select:
346 	if (oc->chosen)
347 		put_task_struct(oc->chosen);
348 	get_task_struct(task);
349 	oc->chosen = task;
350 	oc->chosen_points = points;
351 next:
352 	return 0;
353 abort:
354 	if (oc->chosen)
355 		put_task_struct(oc->chosen);
356 	oc->chosen = (void *)-1UL;
357 	return 1;
358 }
359 
360 /*
361  * Simple selection loop. We choose the process with the highest number of
362  * 'points'. In case scan was aborted, oc->chosen is set to -1.
363  */
select_bad_process(struct oom_control * oc)364 static void select_bad_process(struct oom_control *oc)
365 {
366 	oc->chosen_points = LONG_MIN;
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 
dump_task(struct task_struct * p,void * arg)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 		 * All of p's threads have already detached their mm's. There's
397 		 * no need to report them; they can't be oom killed anyway.
398 		 */
399 		return 0;
400 	}
401 
402 	pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu         %5hd %s\n",
403 		task->pid, from_kuid(&init_user_ns, task_uid(task)),
404 		task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
405 		mm_pgtables_bytes(task->mm),
406 		get_mm_counter(task->mm, MM_SWAPENTS),
407 		task->signal->oom_score_adj, task->comm);
408 	task_unlock(task);
409 
410 	return 0;
411 }
412 
413 /**
414  * dump_tasks - dump current memory state of all system tasks
415  * @oc: pointer to struct oom_control
416  *
417  * Dumps the current memory state of all eligible tasks.  Tasks not in the same
418  * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
419  * are not shown.
420  * State information includes task's pid, uid, tgid, vm size, rss,
421  * pgtables_bytes, swapents, oom_score_adj value, and name.
422  */
dump_tasks(struct oom_control * oc)423 static void dump_tasks(struct oom_control *oc)
424 {
425 	pr_info("Tasks state (memory values in pages):\n");
426 	pr_info("[  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name\n");
427 
428 	if (is_memcg_oom(oc))
429 		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
430 	else {
431 		struct task_struct *p;
432 
433 		rcu_read_lock();
434 		for_each_process(p)
435 			dump_task(p, oc);
436 		rcu_read_unlock();
437 	}
438 }
439 
dump_oom_summary(struct oom_control * oc,struct task_struct * victim)440 static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
441 {
442 	/* one line summary of the oom killer context. */
443 	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
444 			oom_constraint_text[oc->constraint],
445 			nodemask_pr_args(oc->nodemask));
446 	cpuset_print_current_mems_allowed();
447 	mem_cgroup_print_oom_context(oc->memcg, victim);
448 	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
449 		from_kuid(&init_user_ns, task_uid(victim)));
450 }
451 
dump_header(struct oom_control * oc,struct task_struct * p)452 static void dump_header(struct oom_control *oc, struct task_struct *p)
453 {
454 	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
455 		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
456 			current->signal->oom_score_adj);
457 	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
458 		pr_warn("COMPACTION is disabled!!!\n");
459 
460 	dump_stack();
461 	if (is_memcg_oom(oc))
462 		mem_cgroup_print_oom_meminfo(oc->memcg);
463 	else {
464 		__show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask, gfp_zone(oc->gfp_mask));
465 		if (should_dump_unreclaim_slab())
466 			dump_unreclaimable_slab();
467 	}
468 	if (sysctl_oom_dump_tasks)
469 		dump_tasks(oc);
470 	if (p)
471 		dump_oom_summary(oc, p);
472 }
473 
474 /*
475  * Number of OOM victims in flight
476  */
477 static atomic_t oom_victims = ATOMIC_INIT(0);
478 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
479 
480 static bool oom_killer_disabled __read_mostly;
481 
482 /*
483  * task->mm can be NULL if the task is the exited group leader.  So to
484  * determine whether the task is using a particular mm, we examine all the
485  * task's threads: if one of those is using this mm then this task was also
486  * using it.
487  */
process_shares_mm(struct task_struct * p,struct mm_struct * mm)488 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
489 {
490 	struct task_struct *t;
491 
492 	for_each_thread(p, t) {
493 		struct mm_struct *t_mm = READ_ONCE(t->mm);
494 		if (t_mm)
495 			return t_mm == mm;
496 	}
497 	return false;
498 }
499 
500 #ifdef CONFIG_MMU
501 /*
502  * OOM Reaper kernel thread which tries to reap the memory used by the OOM
503  * victim (if that is possible) to help the OOM killer to move on.
504  */
505 static struct task_struct *oom_reaper_th;
506 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
507 static struct task_struct *oom_reaper_list;
508 static DEFINE_SPINLOCK(oom_reaper_lock);
509 
__oom_reap_task_mm(struct mm_struct * mm)510 static bool __oom_reap_task_mm(struct mm_struct *mm)
511 {
512 	struct vm_area_struct *vma;
513 	bool ret = true;
514 	VMA_ITERATOR(vmi, mm, 0);
515 
516 	/*
517 	 * Tell all users of get_user/copy_from_user etc... that the content
518 	 * is no longer stable. No barriers really needed because unmapping
519 	 * should imply barriers already and the reader would hit a page fault
520 	 * if it stumbled over a reaped memory.
521 	 */
522 	set_bit(MMF_UNSTABLE, &mm->flags);
523 
524 	for_each_vma(vmi, vma) {
525 		if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
526 			continue;
527 
528 		/*
529 		 * Only anonymous pages have a good chance to be dropped
530 		 * without additional steps which we cannot afford as we
531 		 * are OOM already.
532 		 *
533 		 * We do not even care about fs backed pages because all
534 		 * which are reclaimable have already been reclaimed and
535 		 * we do not want to block exit_mmap by keeping mm ref
536 		 * count elevated without a good reason.
537 		 */
538 		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
539 			struct mmu_notifier_range range;
540 			struct mmu_gather tlb;
541 
542 			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
543 						mm, vma->vm_start,
544 						vma->vm_end);
545 			tlb_gather_mmu(&tlb, mm);
546 			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
547 				tlb_finish_mmu(&tlb);
548 				ret = false;
549 				continue;
550 			}
551 			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
552 			mmu_notifier_invalidate_range_end(&range);
553 			tlb_finish_mmu(&tlb);
554 		}
555 	}
556 
557 	return ret;
558 }
559 
560 /*
561  * Reaps the address space of the give task.
562  *
563  * Returns true on success and false if none or part of the address space
564  * has been reclaimed and the caller should retry later.
565  */
oom_reap_task_mm(struct task_struct * tsk,struct mm_struct * mm)566 static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
567 {
568 	bool ret = true;
569 
570 	if (!mmap_read_trylock(mm)) {
571 		trace_skip_task_reaping(tsk->pid);
572 		return false;
573 	}
574 
575 	/*
576 	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
577 	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
578 	 * under mmap_lock for reading because it serializes against the
579 	 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
580 	 */
581 	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
582 		trace_skip_task_reaping(tsk->pid);
583 		goto out_unlock;
584 	}
585 
586 	trace_start_task_reaping(tsk->pid);
587 
588 	/* failed to reap part of the address space. Try again later */
589 	ret = __oom_reap_task_mm(mm);
590 	if (!ret)
591 		goto out_finish;
592 
593 	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
594 			task_pid_nr(tsk), tsk->comm,
595 			K(get_mm_counter(mm, MM_ANONPAGES)),
596 			K(get_mm_counter(mm, MM_FILEPAGES)),
597 			K(get_mm_counter(mm, MM_SHMEMPAGES)));
598 out_finish:
599 	trace_finish_task_reaping(tsk->pid);
600 out_unlock:
601 	mmap_read_unlock(mm);
602 
603 	return ret;
604 }
605 
606 #define MAX_OOM_REAP_RETRIES 10
oom_reap_task(struct task_struct * tsk)607 static void oom_reap_task(struct task_struct *tsk)
608 {
609 	int attempts = 0;
610 	struct mm_struct *mm = tsk->signal->oom_mm;
611 
612 	/* Retry the mmap_read_trylock(mm) a few times */
613 	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
614 		schedule_timeout_idle(HZ/10);
615 
616 	if (attempts <= MAX_OOM_REAP_RETRIES ||
617 	    test_bit(MMF_OOM_SKIP, &mm->flags))
618 		goto done;
619 
620 	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
621 		task_pid_nr(tsk), tsk->comm);
622 	sched_show_task(tsk);
623 	debug_show_all_locks();
624 
625 done:
626 	tsk->oom_reaper_list = NULL;
627 
628 	/*
629 	 * Hide this mm from OOM killer because it has been either reaped or
630 	 * somebody can't call mmap_write_unlock(mm).
631 	 */
632 	set_bit(MMF_OOM_SKIP, &mm->flags);
633 
634 	/* Drop a reference taken by queue_oom_reaper */
635 	put_task_struct(tsk);
636 }
637 
oom_reaper(void * unused)638 static int oom_reaper(void *unused)
639 {
640 	set_freezable();
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_irq(&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_irq(&oom_reaper_lock);
652 
653 		if (tsk)
654 			oom_reap_task(tsk);
655 	}
656 
657 	return 0;
658 }
659 
wake_oom_reaper(struct timer_list * timer)660 static void wake_oom_reaper(struct timer_list *timer)
661 {
662 	struct task_struct *tsk = container_of(timer, struct task_struct,
663 			oom_reaper_timer);
664 	struct mm_struct *mm = tsk->signal->oom_mm;
665 	unsigned long flags;
666 
667 	/* The victim managed to terminate on its own - see exit_mmap */
668 	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
669 		put_task_struct(tsk);
670 		return;
671 	}
672 
673 	spin_lock_irqsave(&oom_reaper_lock, flags);
674 	tsk->oom_reaper_list = oom_reaper_list;
675 	oom_reaper_list = tsk;
676 	spin_unlock_irqrestore(&oom_reaper_lock, flags);
677 	trace_wake_reaper(tsk->pid);
678 	wake_up(&oom_reaper_wait);
679 }
680 
681 /*
682  * Give the OOM victim time to exit naturally before invoking the oom_reaping.
683  * The timers timeout is arbitrary... the longer it is, the longer the worst
684  * case scenario for the OOM can take. If it is too small, the oom_reaper can
685  * get in the way and release resources needed by the process exit path.
686  * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
687  * before the exit path is able to wake the futex waiters.
688  */
689 #define OOM_REAPER_DELAY (2*HZ)
queue_oom_reaper(struct task_struct * tsk)690 static void queue_oom_reaper(struct task_struct *tsk)
691 {
692 	/* mm is already queued? */
693 	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
694 		return;
695 
696 	get_task_struct(tsk);
697 	timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
698 	tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
699 	add_timer(&tsk->oom_reaper_timer);
700 }
701 
702 #ifdef CONFIG_SYSCTL
703 static struct ctl_table vm_oom_kill_table[] = {
704 	{
705 		.procname	= "panic_on_oom",
706 		.data		= &sysctl_panic_on_oom,
707 		.maxlen		= sizeof(sysctl_panic_on_oom),
708 		.mode		= 0644,
709 		.proc_handler	= proc_dointvec_minmax,
710 		.extra1		= SYSCTL_ZERO,
711 		.extra2		= SYSCTL_TWO,
712 	},
713 	{
714 		.procname	= "oom_kill_allocating_task",
715 		.data		= &sysctl_oom_kill_allocating_task,
716 		.maxlen		= sizeof(sysctl_oom_kill_allocating_task),
717 		.mode		= 0644,
718 		.proc_handler	= proc_dointvec,
719 	},
720 	{
721 		.procname	= "oom_dump_tasks",
722 		.data		= &sysctl_oom_dump_tasks,
723 		.maxlen		= sizeof(sysctl_oom_dump_tasks),
724 		.mode		= 0644,
725 		.proc_handler	= proc_dointvec,
726 	},
727 	{}
728 };
729 #endif
730 
oom_init(void)731 static int __init oom_init(void)
732 {
733 	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
734 #ifdef CONFIG_SYSCTL
735 	register_sysctl_init("vm", vm_oom_kill_table);
736 #endif
737 	return 0;
738 }
subsys_initcall(oom_init)739 subsys_initcall(oom_init)
740 #else
741 static inline void queue_oom_reaper(struct task_struct *tsk)
742 {
743 }
744 #endif /* CONFIG_MMU */
745 
746 /**
747  * mark_oom_victim - mark the given task as OOM victim
748  * @tsk: task to mark
749  *
750  * Has to be called with oom_lock held and never after
751  * oom has been disabled already.
752  *
753  * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
754  * under task_lock or operate on the current).
755  */
756 static void mark_oom_victim(struct task_struct *tsk)
757 {
758 	struct mm_struct *mm = tsk->mm;
759 
760 	WARN_ON(oom_killer_disabled);
761 	/* OOM killer might race with memcg OOM */
762 	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
763 		return;
764 
765 	/* oom_mm is bound to the signal struct life time. */
766 	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
767 		mmgrab(tsk->signal->oom_mm);
768 
769 	/*
770 	 * Make sure that the task is woken up from uninterruptible sleep
771 	 * if it is frozen because OOM killer wouldn't be able to free
772 	 * any memory and livelock. freezing_slow_path will tell the freezer
773 	 * that TIF_MEMDIE tasks should be ignored.
774 	 */
775 	__thaw_task(tsk);
776 	atomic_inc(&oom_victims);
777 	trace_mark_victim(tsk->pid);
778 }
779 
780 /**
781  * exit_oom_victim - note the exit of an OOM victim
782  */
exit_oom_victim(void)783 void exit_oom_victim(void)
784 {
785 	clear_thread_flag(TIF_MEMDIE);
786 
787 	if (!atomic_dec_return(&oom_victims))
788 		wake_up_all(&oom_victims_wait);
789 }
790 
791 /**
792  * oom_killer_enable - enable OOM killer
793  */
oom_killer_enable(void)794 void oom_killer_enable(void)
795 {
796 	oom_killer_disabled = false;
797 	pr_info("OOM killer enabled.\n");
798 }
799 
800 /**
801  * oom_killer_disable - disable OOM killer
802  * @timeout: maximum timeout to wait for oom victims in jiffies
803  *
804  * Forces all page allocations to fail rather than trigger OOM killer.
805  * Will block and wait until all OOM victims are killed or the given
806  * timeout expires.
807  *
808  * The function cannot be called when there are runnable user tasks because
809  * the userspace would see unexpected allocation failures as a result. Any
810  * new usage of this function should be consulted with MM people.
811  *
812  * Returns true if successful and false if the OOM killer cannot be
813  * disabled.
814  */
oom_killer_disable(signed long timeout)815 bool oom_killer_disable(signed long timeout)
816 {
817 	signed long ret;
818 
819 	/*
820 	 * Make sure to not race with an ongoing OOM killer. Check that the
821 	 * current is not killed (possibly due to sharing the victim's memory).
822 	 */
823 	if (mutex_lock_killable(&oom_lock))
824 		return false;
825 	oom_killer_disabled = true;
826 	mutex_unlock(&oom_lock);
827 
828 	ret = wait_event_interruptible_timeout(oom_victims_wait,
829 			!atomic_read(&oom_victims), timeout);
830 	if (ret <= 0) {
831 		oom_killer_enable();
832 		return false;
833 	}
834 	pr_info("OOM killer disabled.\n");
835 
836 	return true;
837 }
838 
__task_will_free_mem(struct task_struct * task)839 static inline bool __task_will_free_mem(struct task_struct *task)
840 {
841 	struct signal_struct *sig = task->signal;
842 
843 	/*
844 	 * A coredumping process may sleep for an extended period in
845 	 * coredump_task_exit(), so the oom killer cannot assume that
846 	 * the process will promptly exit and release memory.
847 	 */
848 	if (sig->core_state)
849 		return false;
850 
851 	if (sig->flags & SIGNAL_GROUP_EXIT)
852 		return true;
853 
854 	if (thread_group_empty(task) && (task->flags & PF_EXITING))
855 		return true;
856 
857 	return false;
858 }
859 
860 /*
861  * Checks whether the given task is dying or exiting and likely to
862  * release its address space. This means that all threads and processes
863  * sharing the same mm have to be killed or exiting.
864  * Caller has to make sure that task->mm is stable (hold task_lock or
865  * it operates on the current).
866  */
task_will_free_mem(struct task_struct * task)867 static bool task_will_free_mem(struct task_struct *task)
868 {
869 	struct mm_struct *mm = task->mm;
870 	struct task_struct *p;
871 	bool ret = true;
872 
873 	/*
874 	 * Skip tasks without mm because it might have passed its exit_mm and
875 	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
876 	 * on that for now. We can consider find_lock_task_mm in future.
877 	 */
878 	if (!mm)
879 		return false;
880 
881 	if (!__task_will_free_mem(task))
882 		return false;
883 
884 	/*
885 	 * This task has already been drained by the oom reaper so there are
886 	 * only small chances it will free some more
887 	 */
888 	if (test_bit(MMF_OOM_SKIP, &mm->flags))
889 		return false;
890 
891 	if (atomic_read(&mm->mm_users) <= 1)
892 		return true;
893 
894 	/*
895 	 * Make sure that all tasks which share the mm with the given tasks
896 	 * are dying as well to make sure that a) nobody pins its mm and
897 	 * b) the task is also reapable by the oom reaper.
898 	 */
899 	rcu_read_lock();
900 	for_each_process(p) {
901 		if (!process_shares_mm(p, mm))
902 			continue;
903 		if (same_thread_group(task, p))
904 			continue;
905 		ret = __task_will_free_mem(p);
906 		if (!ret)
907 			break;
908 	}
909 	rcu_read_unlock();
910 
911 	return ret;
912 }
913 
__oom_kill_process(struct task_struct * victim,const char * message)914 static void __oom_kill_process(struct task_struct *victim, const char *message)
915 {
916 	struct task_struct *p;
917 	struct mm_struct *mm;
918 	bool can_oom_reap = true;
919 
920 	p = find_lock_task_mm(victim);
921 	if (!p) {
922 		pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
923 			message, task_pid_nr(victim), victim->comm);
924 		put_task_struct(victim);
925 		return;
926 	} else if (victim != p) {
927 		get_task_struct(p);
928 		put_task_struct(victim);
929 		victim = p;
930 	}
931 
932 	/* Get a reference to safely compare mm after task_unlock(victim) */
933 	mm = victim->mm;
934 	mmgrab(mm);
935 
936 	/* Raise event before sending signal: task reaper must see this */
937 	count_vm_event(OOM_KILL);
938 	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
939 
940 	/*
941 	 * We should send SIGKILL before granting access to memory reserves
942 	 * in order to prevent the OOM victim from depleting the memory
943 	 * reserves from the user space under its control.
944 	 */
945 	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
946 	mark_oom_victim(victim);
947 	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",
948 		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
949 		K(get_mm_counter(mm, MM_ANONPAGES)),
950 		K(get_mm_counter(mm, MM_FILEPAGES)),
951 		K(get_mm_counter(mm, MM_SHMEMPAGES)),
952 		from_kuid(&init_user_ns, task_uid(victim)),
953 		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
954 	task_unlock(victim);
955 
956 	/*
957 	 * Kill all user processes sharing victim->mm in other thread groups, if
958 	 * any.  They don't get access to memory reserves, though, to avoid
959 	 * depletion of all memory.  This prevents mm->mmap_lock livelock when an
960 	 * oom killed thread cannot exit because it requires the semaphore and
961 	 * its contended by another thread trying to allocate memory itself.
962 	 * That thread will now get access to memory reserves since it has a
963 	 * pending fatal signal.
964 	 */
965 	rcu_read_lock();
966 	for_each_process(p) {
967 		if (!process_shares_mm(p, mm))
968 			continue;
969 		if (same_thread_group(p, victim))
970 			continue;
971 		if (is_global_init(p)) {
972 			can_oom_reap = false;
973 			set_bit(MMF_OOM_SKIP, &mm->flags);
974 			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
975 					task_pid_nr(victim), victim->comm,
976 					task_pid_nr(p), p->comm);
977 			continue;
978 		}
979 		/*
980 		 * No kthread_use_mm() user needs to read from the userspace so
981 		 * we are ok to reap it.
982 		 */
983 		if (unlikely(p->flags & PF_KTHREAD))
984 			continue;
985 		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
986 	}
987 	rcu_read_unlock();
988 
989 	if (can_oom_reap)
990 		queue_oom_reaper(victim);
991 
992 	mmdrop(mm);
993 	put_task_struct(victim);
994 }
995 
996 /*
997  * Kill provided task unless it's secured by setting
998  * oom_score_adj to OOM_SCORE_ADJ_MIN.
999  */
oom_kill_memcg_member(struct task_struct * task,void * message)1000 static int oom_kill_memcg_member(struct task_struct *task, void *message)
1001 {
1002 	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1003 	    !is_global_init(task)) {
1004 		get_task_struct(task);
1005 		__oom_kill_process(task, message);
1006 	}
1007 	return 0;
1008 }
1009 
oom_kill_process(struct oom_control * oc,const char * message)1010 static void oom_kill_process(struct oom_control *oc, const char *message)
1011 {
1012 	struct task_struct *victim = oc->chosen;
1013 	struct mem_cgroup *oom_group;
1014 	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1015 					      DEFAULT_RATELIMIT_BURST);
1016 
1017 	/*
1018 	 * If the task is already exiting, don't alarm the sysadmin or kill
1019 	 * its children or threads, just give it access to memory reserves
1020 	 * so it can die quickly
1021 	 */
1022 	task_lock(victim);
1023 	if (task_will_free_mem(victim)) {
1024 		mark_oom_victim(victim);
1025 		queue_oom_reaper(victim);
1026 		task_unlock(victim);
1027 		put_task_struct(victim);
1028 		return;
1029 	}
1030 	task_unlock(victim);
1031 
1032 	if (__ratelimit(&oom_rs))
1033 		dump_header(oc, victim);
1034 
1035 	/*
1036 	 * Do we need to kill the entire memory cgroup?
1037 	 * Or even one of the ancestor memory cgroups?
1038 	 * Check this out before killing the victim task.
1039 	 */
1040 	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1041 
1042 	__oom_kill_process(victim, message);
1043 
1044 	/*
1045 	 * If necessary, kill all tasks in the selected memory cgroup.
1046 	 */
1047 	if (oom_group) {
1048 		memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1049 		mem_cgroup_print_oom_group(oom_group);
1050 		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1051 				      (void *)message);
1052 		mem_cgroup_put(oom_group);
1053 	}
1054 }
1055 
1056 /*
1057  * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1058  */
check_panic_on_oom(struct oom_control * oc)1059 static void check_panic_on_oom(struct oom_control *oc)
1060 {
1061 	if (likely(!sysctl_panic_on_oom))
1062 		return;
1063 	if (sysctl_panic_on_oom != 2) {
1064 		/*
1065 		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1066 		 * does not panic for cpuset, mempolicy, or memcg allocation
1067 		 * failures.
1068 		 */
1069 		if (oc->constraint != CONSTRAINT_NONE)
1070 			return;
1071 	}
1072 	/* Do not panic for oom kills triggered by sysrq */
1073 	if (is_sysrq_oom(oc))
1074 		return;
1075 	dump_header(oc, NULL);
1076 	panic("Out of memory: %s panic_on_oom is enabled\n",
1077 		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1078 }
1079 
1080 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1081 
register_oom_notifier(struct notifier_block * nb)1082 int register_oom_notifier(struct notifier_block *nb)
1083 {
1084 	return blocking_notifier_chain_register(&oom_notify_list, nb);
1085 }
1086 EXPORT_SYMBOL_GPL(register_oom_notifier);
1087 
unregister_oom_notifier(struct notifier_block * nb)1088 int unregister_oom_notifier(struct notifier_block *nb)
1089 {
1090 	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1091 }
1092 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1093 
1094 /**
1095  * out_of_memory - kill the "best" process when we run out of memory
1096  * @oc: pointer to struct oom_control
1097  *
1098  * If we run out of memory, we have the choice between either
1099  * killing a random task (bad), letting the system crash (worse)
1100  * OR try to be smart about which process to kill. Note that we
1101  * don't have to be perfect here, we just have to be good.
1102  */
out_of_memory(struct oom_control * oc)1103 bool out_of_memory(struct oom_control *oc)
1104 {
1105 	unsigned long freed = 0;
1106 
1107 	if (oom_killer_disabled)
1108 		return false;
1109 
1110 	if (!is_memcg_oom(oc)) {
1111 		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1112 		if (freed > 0 && !is_sysrq_oom(oc))
1113 			/* Got some memory back in the last second. */
1114 			return true;
1115 	}
1116 
1117 	/*
1118 	 * If current has a pending SIGKILL or is exiting, then automatically
1119 	 * select it.  The goal is to allow it to allocate so that it may
1120 	 * quickly exit and free its memory.
1121 	 */
1122 	if (task_will_free_mem(current)) {
1123 		mark_oom_victim(current);
1124 		queue_oom_reaper(current);
1125 		return true;
1126 	}
1127 
1128 	/*
1129 	 * The OOM killer does not compensate for IO-less reclaim.
1130 	 * But mem_cgroup_oom() has to invoke the OOM killer even
1131 	 * if it is a GFP_NOFS allocation.
1132 	 */
1133 	if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1134 		return true;
1135 
1136 	/*
1137 	 * Check if there were limitations on the allocation (only relevant for
1138 	 * NUMA and memcg) that may require different handling.
1139 	 */
1140 	oc->constraint = constrained_alloc(oc);
1141 	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1142 		oc->nodemask = NULL;
1143 	check_panic_on_oom(oc);
1144 
1145 	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1146 	    current->mm && !oom_unkillable_task(current) &&
1147 	    oom_cpuset_eligible(current, oc) &&
1148 	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1149 		get_task_struct(current);
1150 		oc->chosen = current;
1151 		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1152 		return true;
1153 	}
1154 
1155 	select_bad_process(oc);
1156 	/* Found nothing?!?! */
1157 	if (!oc->chosen) {
1158 		dump_header(oc, NULL);
1159 		pr_warn("Out of memory and no killable processes...\n");
1160 		/*
1161 		 * If we got here due to an actual allocation at the
1162 		 * system level, we cannot survive this and will enter
1163 		 * an endless loop in the allocator. Bail out now.
1164 		 */
1165 		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1166 			panic("System is deadlocked on memory\n");
1167 	}
1168 	if (oc->chosen && oc->chosen != (void *)-1UL)
1169 		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1170 				 "Memory cgroup out of memory");
1171 	return !!oc->chosen;
1172 }
1173 
1174 /*
1175  * The pagefault handler calls here because some allocation has failed. We have
1176  * to take care of the memcg OOM here because this is the only safe context without
1177  * any locks held but let the oom killer triggered from the allocation context care
1178  * about the global OOM.
1179  */
pagefault_out_of_memory(void)1180 void pagefault_out_of_memory(void)
1181 {
1182 	static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1183 				      DEFAULT_RATELIMIT_BURST);
1184 
1185 	if (mem_cgroup_oom_synchronize(true))
1186 		return;
1187 
1188 	if (fatal_signal_pending(current))
1189 		return;
1190 
1191 	if (__ratelimit(&pfoom_rs))
1192 		pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1193 }
1194 
SYSCALL_DEFINE2(process_mrelease,int,pidfd,unsigned int,flags)1195 SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1196 {
1197 #ifdef CONFIG_MMU
1198 	struct mm_struct *mm = NULL;
1199 	struct task_struct *task;
1200 	struct task_struct *p;
1201 	unsigned int f_flags;
1202 	bool reap = false;
1203 	long ret = 0;
1204 
1205 	if (flags)
1206 		return -EINVAL;
1207 
1208 	task = pidfd_get_task(pidfd, &f_flags);
1209 	if (IS_ERR(task))
1210 		return PTR_ERR(task);
1211 
1212 	/*
1213 	 * Make sure to choose a thread which still has a reference to mm
1214 	 * during the group exit
1215 	 */
1216 	p = find_lock_task_mm(task);
1217 	if (!p) {
1218 		ret = -ESRCH;
1219 		goto put_task;
1220 	}
1221 
1222 	mm = p->mm;
1223 	mmgrab(mm);
1224 
1225 	if (task_will_free_mem(p))
1226 		reap = true;
1227 	else {
1228 		/* Error only if the work has not been done already */
1229 		if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1230 			ret = -EINVAL;
1231 	}
1232 	task_unlock(p);
1233 
1234 	if (!reap)
1235 		goto drop_mm;
1236 
1237 	if (mmap_read_lock_killable(mm)) {
1238 		ret = -EINTR;
1239 		goto drop_mm;
1240 	}
1241 	/*
1242 	 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1243 	 * possible change in exit_mmap is seen
1244 	 */
1245 	if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1246 		ret = -EAGAIN;
1247 	mmap_read_unlock(mm);
1248 
1249 drop_mm:
1250 	mmdrop(mm);
1251 put_task:
1252 	put_task_struct(task);
1253 	return ret;
1254 #else
1255 	return -ENOSYS;
1256 #endif /* CONFIG_MMU */
1257 }
1258