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 #include <linux/cred.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
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 */
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
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 */
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 */
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. */
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 */
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 */
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 */
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
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 */
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
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 */
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
435 rcu_read_lock();
436 for_each_process(p)
437 dump_task(p, oc);
438 rcu_read_unlock();
439 }
440 }
441
442 static void dump_oom_victim(struct oom_control *oc, struct task_struct *victim)
443 {
444 /* one line summary of the oom killer context. */
445 pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
446 oom_constraint_text[oc->constraint],
447 nodemask_pr_args(oc->nodemask));
448 cpuset_print_current_mems_allowed();
449 mem_cgroup_print_oom_context(oc->memcg, victim);
450 pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
451 from_kuid(&init_user_ns, task_uid(victim)));
452 }
453
454 static void dump_header(struct oom_control *oc)
455 {
456 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
457 current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
458 current->signal->oom_score_adj);
459 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
460 pr_warn("COMPACTION is disabled!!!\n");
461
462 dump_stack();
463 if (is_memcg_oom(oc))
464 mem_cgroup_print_oom_meminfo(oc->memcg);
465 else {
466 __show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask, gfp_zone(oc->gfp_mask));
467 if (should_dump_unreclaim_slab())
468 dump_unreclaimable_slab();
469 }
470 if (sysctl_oom_dump_tasks)
471 dump_tasks(oc);
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 */
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
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 */
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
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
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
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)
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 #endif
729
730 static int __init oom_init(void)
731 {
732 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
733 #ifdef CONFIG_SYSCTL
734 register_sysctl_init("vm", vm_oom_kill_table);
735 #endif
736 return 0;
737 }
738 subsys_initcall(oom_init)
739 #else
740 static inline void queue_oom_reaper(struct task_struct *tsk)
741 {
742 }
743 #endif /* CONFIG_MMU */
744
745 /**
746 * mark_oom_victim - mark the given task as OOM victim
747 * @tsk: task to mark
748 *
749 * Has to be called with oom_lock held and never after
750 * oom has been disabled already.
751 *
752 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
753 * under task_lock or operate on the current).
754 */
755 static void mark_oom_victim(struct task_struct *tsk)
756 {
757 const struct cred *cred;
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 cred = get_task_cred(tsk);
778 trace_mark_victim(tsk, cred->uid.val);
779 put_cred(cred);
780 }
781
782 /**
783 * exit_oom_victim - note the exit of an OOM victim
784 */
785 void exit_oom_victim(void)
786 {
787 clear_thread_flag(TIF_MEMDIE);
788
789 if (!atomic_dec_return(&oom_victims))
790 wake_up_all(&oom_victims_wait);
791 }
792
793 /**
794 * oom_killer_enable - enable OOM killer
795 */
796 void oom_killer_enable(void)
797 {
798 oom_killer_disabled = false;
799 pr_info("OOM killer enabled.\n");
800 }
801
802 /**
803 * oom_killer_disable - disable OOM killer
804 * @timeout: maximum timeout to wait for oom victims in jiffies
805 *
806 * Forces all page allocations to fail rather than trigger OOM killer.
807 * Will block and wait until all OOM victims are killed or the given
808 * timeout expires.
809 *
810 * The function cannot be called when there are runnable user tasks because
811 * the userspace would see unexpected allocation failures as a result. Any
812 * new usage of this function should be consulted with MM people.
813 *
814 * Returns true if successful and false if the OOM killer cannot be
815 * disabled.
816 */
817 bool oom_killer_disable(signed long timeout)
818 {
819 signed long ret;
820
821 /*
822 * Make sure to not race with an ongoing OOM killer. Check that the
823 * current is not killed (possibly due to sharing the victim's memory).
824 */
825 if (mutex_lock_killable(&oom_lock))
826 return false;
827 oom_killer_disabled = true;
828 mutex_unlock(&oom_lock);
829
830 ret = wait_event_interruptible_timeout(oom_victims_wait,
831 !atomic_read(&oom_victims), timeout);
832 if (ret <= 0) {
833 oom_killer_enable();
834 return false;
835 }
836 pr_info("OOM killer disabled.\n");
837
838 return true;
839 }
840
841 static inline bool __task_will_free_mem(struct task_struct *task)
842 {
843 struct signal_struct *sig = task->signal;
844
845 /*
846 * A coredumping process may sleep for an extended period in
847 * coredump_task_exit(), so the oom killer cannot assume that
848 * the process will promptly exit and release memory.
849 */
850 if (sig->core_state)
851 return false;
852
853 if (sig->flags & SIGNAL_GROUP_EXIT)
854 return true;
855
856 if (thread_group_empty(task) && (task->flags & PF_EXITING))
857 return true;
858
859 return false;
860 }
861
862 /*
863 * Checks whether the given task is dying or exiting and likely to
864 * release its address space. This means that all threads and processes
865 * sharing the same mm have to be killed or exiting.
866 * Caller has to make sure that task->mm is stable (hold task_lock or
867 * it operates on the current).
868 */
869 static bool task_will_free_mem(struct task_struct *task)
870 {
871 struct mm_struct *mm = task->mm;
872 struct task_struct *p;
873 bool ret = true;
874
875 /*
876 * Skip tasks without mm because it might have passed its exit_mm and
877 * exit_oom_victim. oom_reaper could have rescued that but do not rely
878 * on that for now. We can consider find_lock_task_mm in future.
879 */
880 if (!mm)
881 return false;
882
883 if (!__task_will_free_mem(task))
884 return false;
885
886 /*
887 * This task has already been drained by the oom reaper so there are
888 * only small chances it will free some more
889 */
890 if (test_bit(MMF_OOM_SKIP, &mm->flags))
891 return false;
892
893 if (atomic_read(&mm->mm_users) <= 1)
894 return true;
895
896 /*
897 * Make sure that all tasks which share the mm with the given tasks
898 * are dying as well to make sure that a) nobody pins its mm and
899 * b) the task is also reapable by the oom reaper.
900 */
901 rcu_read_lock();
902 for_each_process(p) {
903 if (!process_shares_mm(p, mm))
904 continue;
905 if (same_thread_group(task, p))
906 continue;
907 ret = __task_will_free_mem(p);
908 if (!ret)
909 break;
910 }
911 rcu_read_unlock();
912
913 return ret;
914 }
915
916 static void __oom_kill_process(struct task_struct *victim, const char *message)
917 {
918 struct task_struct *p;
919 struct mm_struct *mm;
920 bool can_oom_reap = true;
921
922 p = find_lock_task_mm(victim);
923 if (!p) {
924 pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
925 message, task_pid_nr(victim), victim->comm);
926 put_task_struct(victim);
927 return;
928 } else if (victim != p) {
929 get_task_struct(p);
930 put_task_struct(victim);
931 victim = p;
932 }
933
934 /* Get a reference to safely compare mm after task_unlock(victim) */
935 mm = victim->mm;
936 mmgrab(mm);
937
938 /* Raise event before sending signal: task reaper must see this */
939 count_vm_event(OOM_KILL);
940 memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
941
942 /*
943 * We should send SIGKILL before granting access to memory reserves
944 * in order to prevent the OOM victim from depleting the memory
945 * reserves from the user space under its control.
946 */
947 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
948 mark_oom_victim(victim);
949 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",
950 message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
951 K(get_mm_counter(mm, MM_ANONPAGES)),
952 K(get_mm_counter(mm, MM_FILEPAGES)),
953 K(get_mm_counter(mm, MM_SHMEMPAGES)),
954 from_kuid(&init_user_ns, task_uid(victim)),
955 mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
956 task_unlock(victim);
957
958 /*
959 * Kill all user processes sharing victim->mm in other thread groups, if
960 * any. They don't get access to memory reserves, though, to avoid
961 * depletion of all memory. This prevents mm->mmap_lock livelock when an
962 * oom killed thread cannot exit because it requires the semaphore and
963 * its contended by another thread trying to allocate memory itself.
964 * That thread will now get access to memory reserves since it has a
965 * pending fatal signal.
966 */
967 rcu_read_lock();
968 for_each_process(p) {
969 if (!process_shares_mm(p, mm))
970 continue;
971 if (same_thread_group(p, victim))
972 continue;
973 if (is_global_init(p)) {
974 can_oom_reap = false;
975 set_bit(MMF_OOM_SKIP, &mm->flags);
976 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
977 task_pid_nr(victim), victim->comm,
978 task_pid_nr(p), p->comm);
979 continue;
980 }
981 /*
982 * No kthread_use_mm() user needs to read from the userspace so
983 * we are ok to reap it.
984 */
985 if (unlikely(p->flags & PF_KTHREAD))
986 continue;
987 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
988 }
989 rcu_read_unlock();
990
991 if (can_oom_reap)
992 queue_oom_reaper(victim);
993
994 mmdrop(mm);
995 put_task_struct(victim);
996 }
997
998 /*
999 * Kill provided task unless it's secured by setting
1000 * oom_score_adj to OOM_SCORE_ADJ_MIN.
1001 */
1002 static int oom_kill_memcg_member(struct task_struct *task, void *message)
1003 {
1004 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1005 !is_global_init(task)) {
1006 get_task_struct(task);
1007 __oom_kill_process(task, message);
1008 }
1009 return 0;
1010 }
1011
1012 static void oom_kill_process(struct oom_control *oc, const char *message)
1013 {
1014 struct task_struct *victim = oc->chosen;
1015 struct mem_cgroup *oom_group;
1016 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1017 DEFAULT_RATELIMIT_BURST);
1018
1019 /*
1020 * If the task is already exiting, don't alarm the sysadmin or kill
1021 * its children or threads, just give it access to memory reserves
1022 * so it can die quickly
1023 */
1024 task_lock(victim);
1025 if (task_will_free_mem(victim)) {
1026 mark_oom_victim(victim);
1027 queue_oom_reaper(victim);
1028 task_unlock(victim);
1029 put_task_struct(victim);
1030 return;
1031 }
1032 task_unlock(victim);
1033
1034 if (__ratelimit(&oom_rs)) {
1035 dump_header(oc);
1036 dump_oom_victim(oc, victim);
1037 }
1038
1039 /*
1040 * Do we need to kill the entire memory cgroup?
1041 * Or even one of the ancestor memory cgroups?
1042 * Check this out before killing the victim task.
1043 */
1044 oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1045
1046 __oom_kill_process(victim, message);
1047
1048 /*
1049 * If necessary, kill all tasks in the selected memory cgroup.
1050 */
1051 if (oom_group) {
1052 memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1053 mem_cgroup_print_oom_group(oom_group);
1054 mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1055 (void *)message);
1056 mem_cgroup_put(oom_group);
1057 }
1058 }
1059
1060 /*
1061 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1062 */
1063 static void check_panic_on_oom(struct oom_control *oc)
1064 {
1065 if (likely(!sysctl_panic_on_oom))
1066 return;
1067 if (sysctl_panic_on_oom != 2) {
1068 /*
1069 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1070 * does not panic for cpuset, mempolicy, or memcg allocation
1071 * failures.
1072 */
1073 if (oc->constraint != CONSTRAINT_NONE)
1074 return;
1075 }
1076 /* Do not panic for oom kills triggered by sysrq */
1077 if (is_sysrq_oom(oc))
1078 return;
1079 dump_header(oc);
1080 panic("Out of memory: %s panic_on_oom is enabled\n",
1081 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1082 }
1083
1084 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1085
1086 int register_oom_notifier(struct notifier_block *nb)
1087 {
1088 return blocking_notifier_chain_register(&oom_notify_list, nb);
1089 }
1090 EXPORT_SYMBOL_GPL(register_oom_notifier);
1091
1092 int unregister_oom_notifier(struct notifier_block *nb)
1093 {
1094 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1095 }
1096 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1097
1098 /**
1099 * out_of_memory - kill the "best" process when we run out of memory
1100 * @oc: pointer to struct oom_control
1101 *
1102 * If we run out of memory, we have the choice between either
1103 * killing a random task (bad), letting the system crash (worse)
1104 * OR try to be smart about which process to kill. Note that we
1105 * don't have to be perfect here, we just have to be good.
1106 */
1107 bool out_of_memory(struct oom_control *oc)
1108 {
1109 unsigned long freed = 0;
1110
1111 if (oom_killer_disabled)
1112 return false;
1113
1114 if (!is_memcg_oom(oc)) {
1115 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1116 if (freed > 0 && !is_sysrq_oom(oc))
1117 /* Got some memory back in the last second. */
1118 return true;
1119 }
1120
1121 /*
1122 * If current has a pending SIGKILL or is exiting, then automatically
1123 * select it. The goal is to allow it to allocate so that it may
1124 * quickly exit and free its memory.
1125 */
1126 if (task_will_free_mem(current)) {
1127 mark_oom_victim(current);
1128 queue_oom_reaper(current);
1129 return true;
1130 }
1131
1132 /*
1133 * The OOM killer does not compensate for IO-less reclaim.
1134 * But mem_cgroup_oom() has to invoke the OOM killer even
1135 * if it is a GFP_NOFS allocation.
1136 */
1137 if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1138 return true;
1139
1140 /*
1141 * Check if there were limitations on the allocation (only relevant for
1142 * NUMA and memcg) that may require different handling.
1143 */
1144 oc->constraint = constrained_alloc(oc);
1145 if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1146 oc->nodemask = NULL;
1147 check_panic_on_oom(oc);
1148
1149 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1150 current->mm && !oom_unkillable_task(current) &&
1151 oom_cpuset_eligible(current, oc) &&
1152 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1153 get_task_struct(current);
1154 oc->chosen = current;
1155 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1156 return true;
1157 }
1158
1159 select_bad_process(oc);
1160 /* Found nothing?!?! */
1161 if (!oc->chosen) {
1162 dump_header(oc);
1163 pr_warn("Out of memory and no killable processes...\n");
1164 /*
1165 * If we got here due to an actual allocation at the
1166 * system level, we cannot survive this and will enter
1167 * an endless loop in the allocator. Bail out now.
1168 */
1169 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1170 panic("System is deadlocked on memory\n");
1171 }
1172 if (oc->chosen && oc->chosen != (void *)-1UL)
1173 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1174 "Memory cgroup out of memory");
1175 return !!oc->chosen;
1176 }
1177
1178 /*
1179 * The pagefault handler calls here because some allocation has failed. We have
1180 * to take care of the memcg OOM here because this is the only safe context without
1181 * any locks held but let the oom killer triggered from the allocation context care
1182 * about the global OOM.
1183 */
1184 void pagefault_out_of_memory(void)
1185 {
1186 static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1187 DEFAULT_RATELIMIT_BURST);
1188
1189 if (mem_cgroup_oom_synchronize(true))
1190 return;
1191
1192 if (fatal_signal_pending(current))
1193 return;
1194
1195 if (__ratelimit(&pfoom_rs))
1196 pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1197 }
1198
1199 SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1200 {
1201 #ifdef CONFIG_MMU
1202 struct mm_struct *mm = NULL;
1203 struct task_struct *task;
1204 struct task_struct *p;
1205 unsigned int f_flags;
1206 bool reap = false;
1207 long ret = 0;
1208
1209 if (flags)
1210 return -EINVAL;
1211
1212 task = pidfd_get_task(pidfd, &f_flags);
1213 if (IS_ERR(task))
1214 return PTR_ERR(task);
1215
1216 /*
1217 * Make sure to choose a thread which still has a reference to mm
1218 * during the group exit
1219 */
1220 p = find_lock_task_mm(task);
1221 if (!p) {
1222 ret = -ESRCH;
1223 goto put_task;
1224 }
1225
1226 mm = p->mm;
1227 mmgrab(mm);
1228
1229 if (task_will_free_mem(p))
1230 reap = true;
1231 else {
1232 /* Error only if the work has not been done already */
1233 if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1234 ret = -EINVAL;
1235 }
1236 task_unlock(p);
1237
1238 if (!reap)
1239 goto drop_mm;
1240
1241 if (mmap_read_lock_killable(mm)) {
1242 ret = -EINTR;
1243 goto drop_mm;
1244 }
1245 /*
1246 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1247 * possible change in exit_mmap is seen
1248 */
1249 if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1250 ret = -EAGAIN;
1251 mmap_read_unlock(mm);
1252
1253 drop_mm:
1254 mmdrop(mm);
1255 put_task:
1256 put_task_struct(task);
1257 return ret;
1258 #else
1259 return -ENOSYS;
1260 #endif /* CONFIG_MMU */
1261 }
1262
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.