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/task.h>
28 #include <linux/sched/debug.h>
29 #include <linux/swap.h>
30 #include <linux/syscalls.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 #include <linux/cred.h>
47 #include <linux/nmi.h>
48 
49 #include <asm/tlb.h>
50 #include "internal.h"
51 #include "slab.h"
52 
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/oom.h>
55 
56 static int sysctl_panic_on_oom;
57 static int sysctl_oom_kill_allocating_task;
58 static int sysctl_oom_dump_tasks = 1;
59 
60 /*
61  * Serializes oom killer invocations (out_of_memory()) from all contexts to
62  * prevent from over eager oom killing (e.g. when the oom killer is invoked
63  * from different domains).
64  *
65  * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
66  * and mark_oom_victim
67  */
68 DEFINE_MUTEX(oom_lock);
69 /* Serializes oom_score_adj and oom_score_adj_min updates */
70 DEFINE_MUTEX(oom_adj_mutex);
71 
is_memcg_oom(struct oom_control * oc)72 static inline bool is_memcg_oom(struct oom_control *oc)
73 {
74 	return oc->memcg != NULL;
75 }
76 
77 #ifdef CONFIG_NUMA
78 /**
79  * oom_cpuset_eligible() - check task eligibility for kill
80  * @start: task struct of which task to consider
81  * @oc: pointer to struct oom_control
82  *
83  * Task eligibility is determined by whether or not a candidate task, @tsk,
84  * shares the same mempolicy nodes as current if it is bound by such a policy
85  * and whether or not it has the same set of allowed cpuset nodes.
86  *
87  * This function is assuming oom-killer context and 'current' has triggered
88  * the oom-killer.
89  */
oom_cpuset_eligible(struct task_struct * start,struct oom_control * oc)90 static bool oom_cpuset_eligible(struct task_struct *start,
91 				struct oom_control *oc)
92 {
93 	struct task_struct *tsk;
94 	bool ret = false;
95 	const nodemask_t *mask = oc->nodemask;
96 
97 	rcu_read_lock();
98 	for_each_thread(start, tsk) {
99 		if (mask) {
100 			/*
101 			 * If this is a mempolicy constrained oom, tsk's
102 			 * cpuset is irrelevant.  Only return true if its
103 			 * mempolicy intersects current, otherwise it may be
104 			 * needlessly killed.
105 			 */
106 			ret = mempolicy_in_oom_domain(tsk, mask);
107 		} else {
108 			/*
109 			 * This is not a mempolicy constrained oom, so only
110 			 * check the mems of tsk's cpuset.
111 			 */
112 			ret = cpuset_mems_allowed_intersects(current, tsk);
113 		}
114 		if (ret)
115 			break;
116 	}
117 	rcu_read_unlock();
118 
119 	return ret;
120 }
121 #else
oom_cpuset_eligible(struct task_struct * tsk,struct oom_control * oc)122 static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
123 {
124 	return true;
125 }
126 #endif /* CONFIG_NUMA */
127 
128 /*
129  * The process p may have detached its own ->mm while exiting or through
130  * kthread_use_mm(), but one or more of its subthreads may still have a valid
131  * pointer.  Return p, or any of its subthreads with a valid ->mm, with
132  * task_lock() held.
133  */
find_lock_task_mm(struct task_struct * p)134 struct task_struct *find_lock_task_mm(struct task_struct *p)
135 {
136 	struct task_struct *t;
137 
138 	rcu_read_lock();
139 
140 	for_each_thread(p, t) {
141 		task_lock(t);
142 		if (likely(t->mm))
143 			goto found;
144 		task_unlock(t);
145 	}
146 	t = NULL;
147 found:
148 	rcu_read_unlock();
149 
150 	return t;
151 }
152 
153 /*
154  * order == -1 means the oom kill is required by sysrq, otherwise only
155  * for display purposes.
156  */
is_sysrq_oom(struct oom_control * oc)157 static inline bool is_sysrq_oom(struct oom_control *oc)
158 {
159 	return oc->order == -1;
160 }
161 
162 /* return true if the task is not adequate as candidate victim task. */
oom_unkillable_task(struct task_struct * p)163 static bool oom_unkillable_task(struct task_struct *p)
164 {
165 	if (is_global_init(p))
166 		return true;
167 	if (p->flags & PF_KTHREAD)
168 		return true;
169 	return false;
170 }
171 
172 /*
173  * Check whether unreclaimable slab amount is greater than
174  * all user memory(LRU pages).
175  * dump_unreclaimable_slab() could help in the case that
176  * oom due to too much unreclaimable slab used by kernel.
177 */
should_dump_unreclaim_slab(void)178 static bool should_dump_unreclaim_slab(void)
179 {
180 	unsigned long nr_lru;
181 
182 	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
183 		 global_node_page_state(NR_INACTIVE_ANON) +
184 		 global_node_page_state(NR_ACTIVE_FILE) +
185 		 global_node_page_state(NR_INACTIVE_FILE) +
186 		 global_node_page_state(NR_ISOLATED_ANON) +
187 		 global_node_page_state(NR_ISOLATED_FILE) +
188 		 global_node_page_state(NR_UNEVICTABLE);
189 
190 	return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
191 }
192 
193 /**
194  * oom_badness - heuristic function to determine which candidate task to kill
195  * @p: task struct of which task we should calculate
196  * @totalpages: total present RAM allowed for page allocation
197  *
198  * The heuristic for determining which task to kill is made to be as simple and
199  * predictable as possible.  The goal is to return the highest value for the
200  * task consuming the most memory to avoid subsequent oom failures.
201  */
oom_badness(struct task_struct * p,unsigned long totalpages)202 long oom_badness(struct task_struct *p, unsigned long totalpages)
203 {
204 	long points;
205 	long adj;
206 
207 	if (oom_unkillable_task(p))
208 		return LONG_MIN;
209 
210 	p = find_lock_task_mm(p);
211 	if (!p)
212 		return LONG_MIN;
213 
214 	/*
215 	 * Do not even consider tasks which are explicitly marked oom
216 	 * unkillable or have been already oom reaped or the are in
217 	 * the middle of vfork
218 	 */
219 	adj = (long)p->signal->oom_score_adj;
220 	if (adj == OOM_SCORE_ADJ_MIN ||
221 			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
222 			in_vfork(p)) {
223 		task_unlock(p);
224 		return LONG_MIN;
225 	}
226 
227 	/*
228 	 * The baseline for the badness score is the proportion of RAM that each
229 	 * task's rss, pagetable and swap space use.
230 	 */
231 	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
232 		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
233 	task_unlock(p);
234 
235 	/* Normalize to oom_score_adj units */
236 	adj *= totalpages / 1000;
237 	points += adj;
238 
239 	return points;
240 }
241 
242 static const char * const oom_constraint_text[] = {
243 	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
244 	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
245 	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
246 	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
247 };
248 
249 /*
250  * Determine the type of allocation constraint.
251  */
constrained_alloc(struct oom_control * oc)252 static enum oom_constraint constrained_alloc(struct oom_control *oc)
253 {
254 	struct zone *zone;
255 	struct zoneref *z;
256 	enum zone_type highest_zoneidx = gfp_zone(oc->gfp_mask);
257 	bool cpuset_limited = false;
258 	int nid;
259 
260 	if (is_memcg_oom(oc)) {
261 		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
262 		return CONSTRAINT_MEMCG;
263 	}
264 
265 	/* Default to all available memory */
266 	oc->totalpages = totalram_pages() + total_swap_pages;
267 
268 	if (!IS_ENABLED(CONFIG_NUMA))
269 		return CONSTRAINT_NONE;
270 
271 	if (!oc->zonelist)
272 		return CONSTRAINT_NONE;
273 	/*
274 	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
275 	 * to kill current.We have to random task kill in this case.
276 	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
277 	 */
278 	if (oc->gfp_mask & __GFP_THISNODE)
279 		return CONSTRAINT_NONE;
280 
281 	/*
282 	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
283 	 * the page allocator means a mempolicy is in effect.  Cpuset policy
284 	 * is enforced in get_page_from_freelist().
285 	 */
286 	if (oc->nodemask &&
287 	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
288 		oc->totalpages = total_swap_pages;
289 		for_each_node_mask(nid, *oc->nodemask)
290 			oc->totalpages += node_present_pages(nid);
291 		return CONSTRAINT_MEMORY_POLICY;
292 	}
293 
294 	/* Check this allocation failure is caused by cpuset's wall function */
295 	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
296 			highest_zoneidx, oc->nodemask)
297 		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
298 			cpuset_limited = true;
299 
300 	if (cpuset_limited) {
301 		oc->totalpages = total_swap_pages;
302 		for_each_node_mask(nid, cpuset_current_mems_allowed)
303 			oc->totalpages += node_present_pages(nid);
304 		return CONSTRAINT_CPUSET;
305 	}
306 	return CONSTRAINT_NONE;
307 }
308 
oom_evaluate_task(struct task_struct * task,void * arg)309 static int oom_evaluate_task(struct task_struct *task, void *arg)
310 {
311 	struct oom_control *oc = arg;
312 	long points;
313 
314 	if (oom_unkillable_task(task))
315 		goto next;
316 
317 	/* p may not have freeable memory in nodemask */
318 	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
319 		goto next;
320 
321 	/*
322 	 * This task already has access to memory reserves and is being killed.
323 	 * Don't allow any other task to have access to the reserves unless
324 	 * the task has MMF_OOM_SKIP because chances that it would release
325 	 * any memory is quite low.
326 	 */
327 	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
328 		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
329 			goto next;
330 		goto abort;
331 	}
332 
333 	/*
334 	 * If task is allocating a lot of memory and has been marked to be
335 	 * killed first if it triggers an oom, then select it.
336 	 */
337 	if (oom_task_origin(task)) {
338 		points = LONG_MAX;
339 		goto select;
340 	}
341 
342 	points = oom_badness(task, oc->totalpages);
343 	if (points == LONG_MIN || points < oc->chosen_points)
344 		goto next;
345 
346 select:
347 	if (oc->chosen)
348 		put_task_struct(oc->chosen);
349 	get_task_struct(task);
350 	oc->chosen = task;
351 	oc->chosen_points = points;
352 next:
353 	return 0;
354 abort:
355 	if (oc->chosen)
356 		put_task_struct(oc->chosen);
357 	oc->chosen = (void *)-1UL;
358 	return 1;
359 }
360 
361 /*
362  * Simple selection loop. We choose the process with the highest number of
363  * 'points'. In case scan was aborted, oc->chosen is set to -1.
364  */
select_bad_process(struct oom_control * oc)365 static void select_bad_process(struct oom_control *oc)
366 {
367 	oc->chosen_points = LONG_MIN;
368 
369 	if (is_memcg_oom(oc))
370 		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
371 	else {
372 		struct task_struct *p;
373 
374 		rcu_read_lock();
375 		for_each_process(p)
376 			if (oom_evaluate_task(p, oc))
377 				break;
378 		rcu_read_unlock();
379 	}
380 }
381 
dump_task(struct task_struct * p,void * arg)382 static int dump_task(struct task_struct *p, void *arg)
383 {
384 	struct oom_control *oc = arg;
385 	struct task_struct *task;
386 
387 	if (oom_unkillable_task(p))
388 		return 0;
389 
390 	/* p may not have freeable memory in nodemask */
391 	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
392 		return 0;
393 
394 	task = find_lock_task_mm(p);
395 	if (!task) {
396 		/*
397 		 * All of p's threads have already detached their mm's. There's
398 		 * no need to report them; they can't be oom killed anyway.
399 		 */
400 		return 0;
401 	}
402 
403 	pr_info("[%7d] %5d %5d %8lu %8lu %8lu %8lu %9lu %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 		get_mm_counter(task->mm, MM_ANONPAGES), get_mm_counter(task->mm, MM_FILEPAGES),
407 		get_mm_counter(task->mm, MM_SHMEMPAGES), mm_pgtables_bytes(task->mm),
408 		get_mm_counter(task->mm, MM_SWAPENTS),
409 		task->signal->oom_score_adj, task->comm);
410 	task_unlock(task);
411 
412 	return 0;
413 }
414 
415 /**
416  * dump_tasks - dump current memory state of all system tasks
417  * @oc: pointer to struct oom_control
418  *
419  * Dumps the current memory state of all eligible tasks.  Tasks not in the same
420  * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
421  * are not shown.
422  * State information includes task's pid, uid, tgid, vm size, rss,
423  * pgtables_bytes, swapents, oom_score_adj value, and name.
424  */
dump_tasks(struct oom_control * oc)425 static void dump_tasks(struct oom_control *oc)
426 {
427 	pr_info("Tasks state (memory values in pages):\n");
428 	pr_info("[  pid  ]   uid  tgid total_vm      rss rss_anon rss_file rss_shmem pgtables_bytes swapents oom_score_adj name\n");
429 
430 	if (is_memcg_oom(oc))
431 		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
432 	else {
433 		struct task_struct *p;
434 		int i = 0;
435 
436 		rcu_read_lock();
437 		for_each_process(p) {
438 			/* Avoid potential softlockup warning */
439 			if ((++i & 1023) == 0)
440 				touch_softlockup_watchdog();
441 			dump_task(p, oc);
442 		}
443 		rcu_read_unlock();
444 	}
445 }
446 
dump_oom_victim(struct oom_control * oc,struct task_struct * victim)447 static void dump_oom_victim(struct oom_control *oc, struct task_struct *victim)
448 {
449 	/* one line summary of the oom killer context. */
450 	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
451 			oom_constraint_text[oc->constraint],
452 			nodemask_pr_args(oc->nodemask));
453 	cpuset_print_current_mems_allowed();
454 	mem_cgroup_print_oom_context(oc->memcg, victim);
455 	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
456 		from_kuid(&init_user_ns, task_uid(victim)));
457 }
458 
dump_header(struct oom_control * oc)459 static void dump_header(struct oom_control *oc)
460 {
461 	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
462 		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
463 			current->signal->oom_score_adj);
464 	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
465 		pr_warn("COMPACTION is disabled!!!\n");
466 
467 	dump_stack();
468 	if (is_memcg_oom(oc))
469 		mem_cgroup_print_oom_meminfo(oc->memcg);
470 	else {
471 		__show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask, gfp_zone(oc->gfp_mask));
472 		if (should_dump_unreclaim_slab())
473 			dump_unreclaimable_slab();
474 	}
475 	if (sysctl_oom_dump_tasks)
476 		dump_tasks(oc);
477 }
478 
479 /*
480  * Number of OOM victims in flight
481  */
482 static atomic_t oom_victims = ATOMIC_INIT(0);
483 static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
484 
485 static bool oom_killer_disabled __read_mostly;
486 
487 /*
488  * task->mm can be NULL if the task is the exited group leader.  So to
489  * determine whether the task is using a particular mm, we examine all the
490  * task's threads: if one of those is using this mm then this task was also
491  * using it.
492  */
process_shares_mm(struct task_struct * p,struct mm_struct * mm)493 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
494 {
495 	struct task_struct *t;
496 
497 	for_each_thread(p, t) {
498 		struct mm_struct *t_mm = READ_ONCE(t->mm);
499 		if (t_mm)
500 			return t_mm == mm;
501 	}
502 	return false;
503 }
504 
505 #ifdef CONFIG_MMU
506 /*
507  * OOM Reaper kernel thread which tries to reap the memory used by the OOM
508  * victim (if that is possible) to help the OOM killer to move on.
509  */
510 static struct task_struct *oom_reaper_th;
511 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
512 static struct task_struct *oom_reaper_list;
513 static DEFINE_SPINLOCK(oom_reaper_lock);
514 
__oom_reap_task_mm(struct mm_struct * mm)515 static bool __oom_reap_task_mm(struct mm_struct *mm)
516 {
517 	struct vm_area_struct *vma;
518 	bool ret = true;
519 	VMA_ITERATOR(vmi, mm, 0);
520 
521 	/*
522 	 * Tell all users of get_user/copy_from_user etc... that the content
523 	 * is no longer stable. No barriers really needed because unmapping
524 	 * should imply barriers already and the reader would hit a page fault
525 	 * if it stumbled over a reaped memory.
526 	 */
527 	set_bit(MMF_UNSTABLE, &mm->flags);
528 
529 	for_each_vma(vmi, vma) {
530 		if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
531 			continue;
532 
533 		/*
534 		 * Only anonymous pages have a good chance to be dropped
535 		 * without additional steps which we cannot afford as we
536 		 * are OOM already.
537 		 *
538 		 * We do not even care about fs backed pages because all
539 		 * which are reclaimable have already been reclaimed and
540 		 * we do not want to block exit_mmap by keeping mm ref
541 		 * count elevated without a good reason.
542 		 */
543 		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
544 			struct mmu_notifier_range range;
545 			struct mmu_gather tlb;
546 
547 			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
548 						mm, vma->vm_start,
549 						vma->vm_end);
550 			tlb_gather_mmu(&tlb, mm);
551 			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
552 				tlb_finish_mmu(&tlb);
553 				ret = false;
554 				continue;
555 			}
556 			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
557 			mmu_notifier_invalidate_range_end(&range);
558 			tlb_finish_mmu(&tlb);
559 		}
560 	}
561 
562 	return ret;
563 }
564 
565 /*
566  * Reaps the address space of the give task.
567  *
568  * Returns true on success and false if none or part of the address space
569  * has been reclaimed and the caller should retry later.
570  */
oom_reap_task_mm(struct task_struct * tsk,struct mm_struct * mm)571 static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
572 {
573 	bool ret = true;
574 
575 	if (!mmap_read_trylock(mm)) {
576 		trace_skip_task_reaping(tsk->pid);
577 		return false;
578 	}
579 
580 	/*
581 	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
582 	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
583 	 * under mmap_lock for reading because it serializes against the
584 	 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
585 	 */
586 	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
587 		trace_skip_task_reaping(tsk->pid);
588 		goto out_unlock;
589 	}
590 
591 	trace_start_task_reaping(tsk->pid);
592 
593 	/* failed to reap part of the address space. Try again later */
594 	ret = __oom_reap_task_mm(mm);
595 	if (!ret)
596 		goto out_finish;
597 
598 	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
599 			task_pid_nr(tsk), tsk->comm,
600 			K(get_mm_counter(mm, MM_ANONPAGES)),
601 			K(get_mm_counter(mm, MM_FILEPAGES)),
602 			K(get_mm_counter(mm, MM_SHMEMPAGES)));
603 out_finish:
604 	trace_finish_task_reaping(tsk->pid);
605 out_unlock:
606 	mmap_read_unlock(mm);
607 
608 	return ret;
609 }
610 
611 #define MAX_OOM_REAP_RETRIES 10
oom_reap_task(struct task_struct * tsk)612 static void oom_reap_task(struct task_struct *tsk)
613 {
614 	int attempts = 0;
615 	struct mm_struct *mm = tsk->signal->oom_mm;
616 
617 	/* Retry the mmap_read_trylock(mm) a few times */
618 	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
619 		schedule_timeout_idle(HZ/10);
620 
621 	if (attempts <= MAX_OOM_REAP_RETRIES ||
622 	    test_bit(MMF_OOM_SKIP, &mm->flags))
623 		goto done;
624 
625 	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
626 		task_pid_nr(tsk), tsk->comm);
627 	sched_show_task(tsk);
628 	debug_show_all_locks();
629 
630 done:
631 	tsk->oom_reaper_list = NULL;
632 
633 	/*
634 	 * Hide this mm from OOM killer because it has been either reaped or
635 	 * somebody can't call mmap_write_unlock(mm).
636 	 */
637 	set_bit(MMF_OOM_SKIP, &mm->flags);
638 
639 	/* Drop a reference taken by queue_oom_reaper */
640 	put_task_struct(tsk);
641 }
642 
oom_reaper(void * unused)643 static int oom_reaper(void *unused)
644 {
645 	set_freezable();
646 
647 	while (true) {
648 		struct task_struct *tsk = NULL;
649 
650 		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
651 		spin_lock_irq(&oom_reaper_lock);
652 		if (oom_reaper_list != NULL) {
653 			tsk = oom_reaper_list;
654 			oom_reaper_list = tsk->oom_reaper_list;
655 		}
656 		spin_unlock_irq(&oom_reaper_lock);
657 
658 		if (tsk)
659 			oom_reap_task(tsk);
660 	}
661 
662 	return 0;
663 }
664 
wake_oom_reaper(struct timer_list * timer)665 static void wake_oom_reaper(struct timer_list *timer)
666 {
667 	struct task_struct *tsk = container_of(timer, struct task_struct,
668 			oom_reaper_timer);
669 	struct mm_struct *mm = tsk->signal->oom_mm;
670 	unsigned long flags;
671 
672 	/* The victim managed to terminate on its own - see exit_mmap */
673 	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
674 		put_task_struct(tsk);
675 		return;
676 	}
677 
678 	spin_lock_irqsave(&oom_reaper_lock, flags);
679 	tsk->oom_reaper_list = oom_reaper_list;
680 	oom_reaper_list = tsk;
681 	spin_unlock_irqrestore(&oom_reaper_lock, flags);
682 	trace_wake_reaper(tsk->pid);
683 	wake_up(&oom_reaper_wait);
684 }
685 
686 /*
687  * Give the OOM victim time to exit naturally before invoking the oom_reaping.
688  * The timers timeout is arbitrary... the longer it is, the longer the worst
689  * case scenario for the OOM can take. If it is too small, the oom_reaper can
690  * get in the way and release resources needed by the process exit path.
691  * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
692  * before the exit path is able to wake the futex waiters.
693  */
694 #define OOM_REAPER_DELAY (2*HZ)
queue_oom_reaper(struct task_struct * tsk)695 static void queue_oom_reaper(struct task_struct *tsk)
696 {
697 	/* mm is already queued? */
698 	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
699 		return;
700 
701 	get_task_struct(tsk);
702 	timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
703 	tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
704 	add_timer(&tsk->oom_reaper_timer);
705 }
706 
707 #ifdef CONFIG_SYSCTL
708 static const struct ctl_table vm_oom_kill_table[] = {
709 	{
710 		.procname	= "panic_on_oom",
711 		.data		= &sysctl_panic_on_oom,
712 		.maxlen		= sizeof(sysctl_panic_on_oom),
713 		.mode		= 0644,
714 		.proc_handler	= proc_dointvec_minmax,
715 		.extra1		= SYSCTL_ZERO,
716 		.extra2		= SYSCTL_TWO,
717 	},
718 	{
719 		.procname	= "oom_kill_allocating_task",
720 		.data		= &sysctl_oom_kill_allocating_task,
721 		.maxlen		= sizeof(sysctl_oom_kill_allocating_task),
722 		.mode		= 0644,
723 		.proc_handler	= proc_dointvec,
724 	},
725 	{
726 		.procname	= "oom_dump_tasks",
727 		.data		= &sysctl_oom_dump_tasks,
728 		.maxlen		= sizeof(sysctl_oom_dump_tasks),
729 		.mode		= 0644,
730 		.proc_handler	= proc_dointvec,
731 	},
732 };
733 #endif
734 
oom_init(void)735 static int __init oom_init(void)
736 {
737 	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
738 #ifdef CONFIG_SYSCTL
739 	register_sysctl_init("vm", vm_oom_kill_table);
740 #endif
741 	return 0;
742 }
subsys_initcall(oom_init)743 subsys_initcall(oom_init)
744 #else
745 static inline void queue_oom_reaper(struct task_struct *tsk)
746 {
747 }
748 #endif /* CONFIG_MMU */
749 
750 /**
751  * mark_oom_victim - mark the given task as OOM victim
752  * @tsk: task to mark
753  *
754  * Has to be called with oom_lock held and never after
755  * oom has been disabled already.
756  *
757  * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
758  * under task_lock or operate on the current).
759  */
760 static void mark_oom_victim(struct task_struct *tsk)
761 {
762 	const struct cred *cred;
763 	struct mm_struct *mm = tsk->mm;
764 
765 	WARN_ON(oom_killer_disabled);
766 	/* OOM killer might race with memcg OOM */
767 	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
768 		return;
769 
770 	/* oom_mm is bound to the signal struct life time. */
771 	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
772 		mmgrab(tsk->signal->oom_mm);
773 
774 	/*
775 	 * Make sure that the task is woken up from uninterruptible sleep
776 	 * if it is frozen because OOM killer wouldn't be able to free
777 	 * any memory and livelock. freezing_slow_path will tell the freezer
778 	 * that TIF_MEMDIE tasks should be ignored.
779 	 */
780 	__thaw_task(tsk);
781 	atomic_inc(&oom_victims);
782 	cred = get_task_cred(tsk);
783 	trace_mark_victim(tsk, cred->uid.val);
784 	put_cred(cred);
785 }
786 
787 /**
788  * exit_oom_victim - note the exit of an OOM victim
789  */
exit_oom_victim(void)790 void exit_oom_victim(void)
791 {
792 	clear_thread_flag(TIF_MEMDIE);
793 
794 	if (!atomic_dec_return(&oom_victims))
795 		wake_up_all(&oom_victims_wait);
796 }
797 
798 /**
799  * oom_killer_enable - enable OOM killer
800  */
oom_killer_enable(void)801 void oom_killer_enable(void)
802 {
803 	oom_killer_disabled = false;
804 	pr_info("OOM killer enabled.\n");
805 }
806 
807 /**
808  * oom_killer_disable - disable OOM killer
809  * @timeout: maximum timeout to wait for oom victims in jiffies
810  *
811  * Forces all page allocations to fail rather than trigger OOM killer.
812  * Will block and wait until all OOM victims are killed or the given
813  * timeout expires.
814  *
815  * The function cannot be called when there are runnable user tasks because
816  * the userspace would see unexpected allocation failures as a result. Any
817  * new usage of this function should be consulted with MM people.
818  *
819  * Returns true if successful and false if the OOM killer cannot be
820  * disabled.
821  */
oom_killer_disable(signed long timeout)822 bool oom_killer_disable(signed long timeout)
823 {
824 	signed long ret;
825 
826 	/*
827 	 * Make sure to not race with an ongoing OOM killer. Check that the
828 	 * current is not killed (possibly due to sharing the victim's memory).
829 	 */
830 	if (mutex_lock_killable(&oom_lock))
831 		return false;
832 	oom_killer_disabled = true;
833 	mutex_unlock(&oom_lock);
834 
835 	ret = wait_event_interruptible_timeout(oom_victims_wait,
836 			!atomic_read(&oom_victims), timeout);
837 	if (ret <= 0) {
838 		oom_killer_enable();
839 		return false;
840 	}
841 	pr_info("OOM killer disabled.\n");
842 
843 	return true;
844 }
845 
__task_will_free_mem(struct task_struct * task)846 static inline bool __task_will_free_mem(struct task_struct *task)
847 {
848 	struct signal_struct *sig = task->signal;
849 
850 	/*
851 	 * A coredumping process may sleep for an extended period in
852 	 * coredump_task_exit(), so the oom killer cannot assume that
853 	 * the process will promptly exit and release memory.
854 	 */
855 	if (sig->core_state)
856 		return false;
857 
858 	if (sig->flags & SIGNAL_GROUP_EXIT)
859 		return true;
860 
861 	if (thread_group_empty(task) && (task->flags & PF_EXITING))
862 		return true;
863 
864 	return false;
865 }
866 
867 /*
868  * Checks whether the given task is dying or exiting and likely to
869  * release its address space. This means that all threads and processes
870  * sharing the same mm have to be killed or exiting.
871  * Caller has to make sure that task->mm is stable (hold task_lock or
872  * it operates on the current).
873  */
task_will_free_mem(struct task_struct * task)874 static bool task_will_free_mem(struct task_struct *task)
875 {
876 	struct mm_struct *mm = task->mm;
877 	struct task_struct *p;
878 	bool ret = true;
879 
880 	/*
881 	 * Skip tasks without mm because it might have passed its exit_mm and
882 	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
883 	 * on that for now. We can consider find_lock_task_mm in future.
884 	 */
885 	if (!mm)
886 		return false;
887 
888 	if (!__task_will_free_mem(task))
889 		return false;
890 
891 	/*
892 	 * This task has already been drained by the oom reaper so there are
893 	 * only small chances it will free some more
894 	 */
895 	if (test_bit(MMF_OOM_SKIP, &mm->flags))
896 		return false;
897 
898 	if (atomic_read(&mm->mm_users) <= 1)
899 		return true;
900 
901 	/*
902 	 * Make sure that all tasks which share the mm with the given tasks
903 	 * are dying as well to make sure that a) nobody pins its mm and
904 	 * b) the task is also reapable by the oom reaper.
905 	 */
906 	rcu_read_lock();
907 	for_each_process(p) {
908 		if (!process_shares_mm(p, mm))
909 			continue;
910 		if (same_thread_group(task, p))
911 			continue;
912 		ret = __task_will_free_mem(p);
913 		if (!ret)
914 			break;
915 	}
916 	rcu_read_unlock();
917 
918 	return ret;
919 }
920 
__oom_kill_process(struct task_struct * victim,const char * message)921 static void __oom_kill_process(struct task_struct *victim, const char *message)
922 {
923 	struct task_struct *p;
924 	struct mm_struct *mm;
925 	bool can_oom_reap = true;
926 
927 	p = find_lock_task_mm(victim);
928 	if (!p) {
929 		pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
930 			message, task_pid_nr(victim), victim->comm);
931 		put_task_struct(victim);
932 		return;
933 	} else if (victim != p) {
934 		get_task_struct(p);
935 		put_task_struct(victim);
936 		victim = p;
937 	}
938 
939 	/* Get a reference to safely compare mm after task_unlock(victim) */
940 	mm = victim->mm;
941 	mmgrab(mm);
942 
943 	/* Raise event before sending signal: task reaper must see this */
944 	count_vm_event(OOM_KILL);
945 	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
946 
947 	/*
948 	 * We should send SIGKILL before granting access to memory reserves
949 	 * in order to prevent the OOM victim from depleting the memory
950 	 * reserves from the user space under its control.
951 	 */
952 	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
953 	mark_oom_victim(victim);
954 	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",
955 		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
956 		K(get_mm_counter(mm, MM_ANONPAGES)),
957 		K(get_mm_counter(mm, MM_FILEPAGES)),
958 		K(get_mm_counter(mm, MM_SHMEMPAGES)),
959 		from_kuid(&init_user_ns, task_uid(victim)),
960 		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
961 	task_unlock(victim);
962 
963 	/*
964 	 * Kill all user processes sharing victim->mm in other thread groups, if
965 	 * any.  They don't get access to memory reserves, though, to avoid
966 	 * depletion of all memory.  This prevents mm->mmap_lock livelock when an
967 	 * oom killed thread cannot exit because it requires the semaphore and
968 	 * its contended by another thread trying to allocate memory itself.
969 	 * That thread will now get access to memory reserves since it has a
970 	 * pending fatal signal.
971 	 */
972 	rcu_read_lock();
973 	for_each_process(p) {
974 		if (!process_shares_mm(p, mm))
975 			continue;
976 		if (same_thread_group(p, victim))
977 			continue;
978 		if (is_global_init(p)) {
979 			can_oom_reap = false;
980 			set_bit(MMF_OOM_SKIP, &mm->flags);
981 			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
982 					task_pid_nr(victim), victim->comm,
983 					task_pid_nr(p), p->comm);
984 			continue;
985 		}
986 		/*
987 		 * No kthread_use_mm() user needs to read from the userspace so
988 		 * we are ok to reap it.
989 		 */
990 		if (unlikely(p->flags & PF_KTHREAD))
991 			continue;
992 		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
993 	}
994 	rcu_read_unlock();
995 
996 	if (can_oom_reap)
997 		queue_oom_reaper(victim);
998 
999 	mmdrop(mm);
1000 	put_task_struct(victim);
1001 }
1002 
1003 /*
1004  * Kill provided task unless it's secured by setting
1005  * oom_score_adj to OOM_SCORE_ADJ_MIN.
1006  */
oom_kill_memcg_member(struct task_struct * task,void * message)1007 static int oom_kill_memcg_member(struct task_struct *task, void *message)
1008 {
1009 	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1010 	    !is_global_init(task)) {
1011 		get_task_struct(task);
1012 		__oom_kill_process(task, message);
1013 	}
1014 	return 0;
1015 }
1016 
oom_kill_process(struct oom_control * oc,const char * message)1017 static void oom_kill_process(struct oom_control *oc, const char *message)
1018 {
1019 	struct task_struct *victim = oc->chosen;
1020 	struct mem_cgroup *oom_group;
1021 	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1022 					      DEFAULT_RATELIMIT_BURST);
1023 
1024 	/*
1025 	 * If the task is already exiting, don't alarm the sysadmin or kill
1026 	 * its children or threads, just give it access to memory reserves
1027 	 * so it can die quickly
1028 	 */
1029 	task_lock(victim);
1030 	if (task_will_free_mem(victim)) {
1031 		mark_oom_victim(victim);
1032 		queue_oom_reaper(victim);
1033 		task_unlock(victim);
1034 		put_task_struct(victim);
1035 		return;
1036 	}
1037 	task_unlock(victim);
1038 
1039 	if (__ratelimit(&oom_rs)) {
1040 		dump_header(oc);
1041 		dump_oom_victim(oc, victim);
1042 	}
1043 
1044 	/*
1045 	 * Do we need to kill the entire memory cgroup?
1046 	 * Or even one of the ancestor memory cgroups?
1047 	 * Check this out before killing the victim task.
1048 	 */
1049 	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1050 
1051 	__oom_kill_process(victim, message);
1052 
1053 	/*
1054 	 * If necessary, kill all tasks in the selected memory cgroup.
1055 	 */
1056 	if (oom_group) {
1057 		memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1058 		mem_cgroup_print_oom_group(oom_group);
1059 		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1060 				      (void *)message);
1061 		mem_cgroup_put(oom_group);
1062 	}
1063 }
1064 
1065 /*
1066  * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1067  */
check_panic_on_oom(struct oom_control * oc)1068 static void check_panic_on_oom(struct oom_control *oc)
1069 {
1070 	if (likely(!sysctl_panic_on_oom))
1071 		return;
1072 	if (sysctl_panic_on_oom != 2) {
1073 		/*
1074 		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1075 		 * does not panic for cpuset, mempolicy, or memcg allocation
1076 		 * failures.
1077 		 */
1078 		if (oc->constraint != CONSTRAINT_NONE)
1079 			return;
1080 	}
1081 	/* Do not panic for oom kills triggered by sysrq */
1082 	if (is_sysrq_oom(oc))
1083 		return;
1084 	dump_header(oc);
1085 	panic("Out of memory: %s panic_on_oom is enabled\n",
1086 		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1087 }
1088 
1089 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1090 
register_oom_notifier(struct notifier_block * nb)1091 int register_oom_notifier(struct notifier_block *nb)
1092 {
1093 	return blocking_notifier_chain_register(&oom_notify_list, nb);
1094 }
1095 EXPORT_SYMBOL_GPL(register_oom_notifier);
1096 
unregister_oom_notifier(struct notifier_block * nb)1097 int unregister_oom_notifier(struct notifier_block *nb)
1098 {
1099 	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1100 }
1101 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1102 
1103 /**
1104  * out_of_memory - kill the "best" process when we run out of memory
1105  * @oc: pointer to struct oom_control
1106  *
1107  * If we run out of memory, we have the choice between either
1108  * killing a random task (bad), letting the system crash (worse)
1109  * OR try to be smart about which process to kill. Note that we
1110  * don't have to be perfect here, we just have to be good.
1111  */
out_of_memory(struct oom_control * oc)1112 bool out_of_memory(struct oom_control *oc)
1113 {
1114 	unsigned long freed = 0;
1115 
1116 	if (oom_killer_disabled)
1117 		return false;
1118 
1119 	if (!is_memcg_oom(oc)) {
1120 		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1121 		if (freed > 0 && !is_sysrq_oom(oc))
1122 			/* Got some memory back in the last second. */
1123 			return true;
1124 	}
1125 
1126 	/*
1127 	 * If current has a pending SIGKILL or is exiting, then automatically
1128 	 * select it.  The goal is to allow it to allocate so that it may
1129 	 * quickly exit and free its memory.
1130 	 */
1131 	if (task_will_free_mem(current)) {
1132 		mark_oom_victim(current);
1133 		queue_oom_reaper(current);
1134 		return true;
1135 	}
1136 
1137 	/*
1138 	 * The OOM killer does not compensate for IO-less reclaim.
1139 	 * But mem_cgroup_oom() has to invoke the OOM killer even
1140 	 * if it is a GFP_NOFS allocation.
1141 	 */
1142 	if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1143 		return true;
1144 
1145 	/*
1146 	 * Check if there were limitations on the allocation (only relevant for
1147 	 * NUMA and memcg) that may require different handling.
1148 	 */
1149 	oc->constraint = constrained_alloc(oc);
1150 	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1151 		oc->nodemask = NULL;
1152 	check_panic_on_oom(oc);
1153 
1154 	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1155 	    current->mm && !oom_unkillable_task(current) &&
1156 	    oom_cpuset_eligible(current, oc) &&
1157 	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1158 		get_task_struct(current);
1159 		oc->chosen = current;
1160 		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1161 		return true;
1162 	}
1163 
1164 	select_bad_process(oc);
1165 	/* Found nothing?!?! */
1166 	if (!oc->chosen) {
1167 		dump_header(oc);
1168 		pr_warn("Out of memory and no killable processes...\n");
1169 		/*
1170 		 * If we got here due to an actual allocation at the
1171 		 * system level, we cannot survive this and will enter
1172 		 * an endless loop in the allocator. Bail out now.
1173 		 */
1174 		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1175 			panic("System is deadlocked on memory\n");
1176 	}
1177 	if (oc->chosen && oc->chosen != (void *)-1UL)
1178 		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1179 				 "Memory cgroup out of memory");
1180 	return !!oc->chosen;
1181 }
1182 
1183 /*
1184  * The pagefault handler calls here because some allocation has failed. We have
1185  * to take care of the memcg OOM here because this is the only safe context without
1186  * any locks held but let the oom killer triggered from the allocation context care
1187  * about the global OOM.
1188  */
pagefault_out_of_memory(void)1189 void pagefault_out_of_memory(void)
1190 {
1191 	static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1192 				      DEFAULT_RATELIMIT_BURST);
1193 
1194 	if (mem_cgroup_oom_synchronize(true))
1195 		return;
1196 
1197 	if (fatal_signal_pending(current))
1198 		return;
1199 
1200 	if (__ratelimit(&pfoom_rs))
1201 		pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1202 }
1203 
SYSCALL_DEFINE2(process_mrelease,int,pidfd,unsigned int,flags)1204 SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1205 {
1206 #ifdef CONFIG_MMU
1207 	struct mm_struct *mm = NULL;
1208 	struct task_struct *task;
1209 	struct task_struct *p;
1210 	unsigned int f_flags;
1211 	bool reap = false;
1212 	long ret = 0;
1213 
1214 	if (flags)
1215 		return -EINVAL;
1216 
1217 	task = pidfd_get_task(pidfd, &f_flags);
1218 	if (IS_ERR(task))
1219 		return PTR_ERR(task);
1220 
1221 	/*
1222 	 * Make sure to choose a thread which still has a reference to mm
1223 	 * during the group exit
1224 	 */
1225 	p = find_lock_task_mm(task);
1226 	if (!p) {
1227 		ret = -ESRCH;
1228 		goto put_task;
1229 	}
1230 
1231 	mm = p->mm;
1232 	mmgrab(mm);
1233 
1234 	if (task_will_free_mem(p))
1235 		reap = true;
1236 	else {
1237 		/* Error only if the work has not been done already */
1238 		if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1239 			ret = -EINVAL;
1240 	}
1241 	task_unlock(p);
1242 
1243 	if (!reap)
1244 		goto drop_mm;
1245 
1246 	if (mmap_read_lock_killable(mm)) {
1247 		ret = -EINTR;
1248 		goto drop_mm;
1249 	}
1250 	/*
1251 	 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1252 	 * possible change in exit_mmap is seen
1253 	 */
1254 	if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1255 		ret = -EAGAIN;
1256 	mmap_read_unlock(mm);
1257 
1258 drop_mm:
1259 	mmdrop(mm);
1260 put_task:
1261 	put_task_struct(task);
1262 	return ret;
1263 #else
1264 	return -ENOSYS;
1265 #endif /* CONFIG_MMU */
1266 }
1267