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 #define K(x) ((x) << (PAGE_SHIFT-10))
483
484 /*
485 * task->mm can be NULL if the task is the exited group leader. So to
486 * determine whether the task is using a particular mm, we examine all the
487 * task's threads: if one of those is using this mm then this task was also
488 * using it.
489 */
process_shares_mm(struct task_struct * p,struct mm_struct * mm)490 bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
491 {
492 struct task_struct *t;
493
494 for_each_thread(p, t) {
495 struct mm_struct *t_mm = READ_ONCE(t->mm);
496 if (t_mm)
497 return t_mm == mm;
498 }
499 return false;
500 }
501
502 #ifdef CONFIG_MMU
503 /*
504 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
505 * victim (if that is possible) to help the OOM killer to move on.
506 */
507 static struct task_struct *oom_reaper_th;
508 static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
509 static struct task_struct *oom_reaper_list;
510 static DEFINE_SPINLOCK(oom_reaper_lock);
511
__oom_reap_task_mm(struct mm_struct * mm)512 static bool __oom_reap_task_mm(struct mm_struct *mm)
513 {
514 struct vm_area_struct *vma;
515 bool ret = true;
516 VMA_ITERATOR(vmi, mm, 0);
517
518 /*
519 * Tell all users of get_user/copy_from_user etc... that the content
520 * is no longer stable. No barriers really needed because unmapping
521 * should imply barriers already and the reader would hit a page fault
522 * if it stumbled over a reaped memory.
523 */
524 set_bit(MMF_UNSTABLE, &mm->flags);
525
526 for_each_vma(vmi, vma) {
527 if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
528 continue;
529
530 /*
531 * Only anonymous pages have a good chance to be dropped
532 * without additional steps which we cannot afford as we
533 * are OOM already.
534 *
535 * We do not even care about fs backed pages because all
536 * which are reclaimable have already been reclaimed and
537 * we do not want to block exit_mmap by keeping mm ref
538 * count elevated without a good reason.
539 */
540 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
541 struct mmu_notifier_range range;
542 struct mmu_gather tlb;
543
544 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
545 vma, mm, vma->vm_start,
546 vma->vm_end);
547 tlb_gather_mmu(&tlb, mm);
548 if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
549 tlb_finish_mmu(&tlb);
550 ret = false;
551 continue;
552 }
553 unmap_page_range(&tlb, vma, range.start, range.end, NULL);
554 mmu_notifier_invalidate_range_end(&range);
555 tlb_finish_mmu(&tlb);
556 }
557 }
558
559 return ret;
560 }
561
562 /*
563 * Reaps the address space of the give task.
564 *
565 * Returns true on success and false if none or part of the address space
566 * has been reclaimed and the caller should retry later.
567 */
oom_reap_task_mm(struct task_struct * tsk,struct mm_struct * mm)568 static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
569 {
570 bool ret = true;
571
572 if (!mmap_read_trylock(mm)) {
573 trace_skip_task_reaping(tsk->pid);
574 return false;
575 }
576
577 /*
578 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
579 * work on the mm anymore. The check for MMF_OOM_SKIP must run
580 * under mmap_lock for reading because it serializes against the
581 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
582 */
583 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
584 trace_skip_task_reaping(tsk->pid);
585 goto out_unlock;
586 }
587
588 trace_start_task_reaping(tsk->pid);
589
590 /* failed to reap part of the address space. Try again later */
591 ret = __oom_reap_task_mm(mm);
592 if (!ret)
593 goto out_finish;
594
595 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
596 task_pid_nr(tsk), tsk->comm,
597 K(get_mm_counter(mm, MM_ANONPAGES)),
598 K(get_mm_counter(mm, MM_FILEPAGES)),
599 K(get_mm_counter(mm, MM_SHMEMPAGES)));
600 out_finish:
601 trace_finish_task_reaping(tsk->pid);
602 out_unlock:
603 mmap_read_unlock(mm);
604
605 return ret;
606 }
607
608 #define MAX_OOM_REAP_RETRIES 10
oom_reap_task(struct task_struct * tsk)609 static void oom_reap_task(struct task_struct *tsk)
610 {
611 int attempts = 0;
612 struct mm_struct *mm = tsk->signal->oom_mm;
613
614 /* Retry the mmap_read_trylock(mm) a few times */
615 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
616 schedule_timeout_idle(HZ/10);
617
618 if (attempts <= MAX_OOM_REAP_RETRIES ||
619 test_bit(MMF_OOM_SKIP, &mm->flags))
620 goto done;
621
622 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
623 task_pid_nr(tsk), tsk->comm);
624 sched_show_task(tsk);
625 debug_show_all_locks();
626
627 done:
628 tsk->oom_reaper_list = NULL;
629
630 /*
631 * Hide this mm from OOM killer because it has been either reaped or
632 * somebody can't call mmap_write_unlock(mm).
633 */
634 set_bit(MMF_OOM_SKIP, &mm->flags);
635
636 /* Drop a reference taken by queue_oom_reaper */
637 put_task_struct(tsk);
638 }
639
oom_reaper(void * unused)640 static int oom_reaper(void *unused)
641 {
642 set_freezable();
643
644 while (true) {
645 struct task_struct *tsk = NULL;
646
647 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
648 spin_lock_irq(&oom_reaper_lock);
649 if (oom_reaper_list != NULL) {
650 tsk = oom_reaper_list;
651 oom_reaper_list = tsk->oom_reaper_list;
652 }
653 spin_unlock_irq(&oom_reaper_lock);
654
655 if (tsk)
656 oom_reap_task(tsk);
657 }
658
659 return 0;
660 }
661
wake_oom_reaper(struct timer_list * timer)662 static void wake_oom_reaper(struct timer_list *timer)
663 {
664 struct task_struct *tsk = container_of(timer, struct task_struct,
665 oom_reaper_timer);
666 struct mm_struct *mm = tsk->signal->oom_mm;
667 unsigned long flags;
668
669 /* The victim managed to terminate on its own - see exit_mmap */
670 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
671 put_task_struct(tsk);
672 return;
673 }
674
675 spin_lock_irqsave(&oom_reaper_lock, flags);
676 tsk->oom_reaper_list = oom_reaper_list;
677 oom_reaper_list = tsk;
678 spin_unlock_irqrestore(&oom_reaper_lock, flags);
679 trace_wake_reaper(tsk->pid);
680 wake_up(&oom_reaper_wait);
681 }
682
683 /*
684 * Give the OOM victim time to exit naturally before invoking the oom_reaping.
685 * The timers timeout is arbitrary... the longer it is, the longer the worst
686 * case scenario for the OOM can take. If it is too small, the oom_reaper can
687 * get in the way and release resources needed by the process exit path.
688 * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
689 * before the exit path is able to wake the futex waiters.
690 */
691 #define OOM_REAPER_DELAY (2*HZ)
queue_oom_reaper(struct task_struct * tsk)692 static void queue_oom_reaper(struct task_struct *tsk)
693 {
694 /* mm is already queued? */
695 if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
696 return;
697
698 get_task_struct(tsk);
699 timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
700 tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
701 add_timer(&tsk->oom_reaper_timer);
702 }
703
704 #ifdef CONFIG_SYSCTL
705 static struct ctl_table vm_oom_kill_table[] = {
706 {
707 .procname = "panic_on_oom",
708 .data = &sysctl_panic_on_oom,
709 .maxlen = sizeof(sysctl_panic_on_oom),
710 .mode = 0644,
711 .proc_handler = proc_dointvec_minmax,
712 .extra1 = SYSCTL_ZERO,
713 .extra2 = SYSCTL_TWO,
714 },
715 {
716 .procname = "oom_kill_allocating_task",
717 .data = &sysctl_oom_kill_allocating_task,
718 .maxlen = sizeof(sysctl_oom_kill_allocating_task),
719 .mode = 0644,
720 .proc_handler = proc_dointvec,
721 },
722 {
723 .procname = "oom_dump_tasks",
724 .data = &sysctl_oom_dump_tasks,
725 .maxlen = sizeof(sysctl_oom_dump_tasks),
726 .mode = 0644,
727 .proc_handler = proc_dointvec,
728 },
729 {}
730 };
731 #endif
732
oom_init(void)733 static int __init oom_init(void)
734 {
735 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
736 #ifdef CONFIG_SYSCTL
737 register_sysctl_init("vm", vm_oom_kill_table);
738 #endif
739 return 0;
740 }
subsys_initcall(oom_init)741 subsys_initcall(oom_init)
742 #else
743 static inline void queue_oom_reaper(struct task_struct *tsk)
744 {
745 }
746 #endif /* CONFIG_MMU */
747
748 /**
749 * mark_oom_victim - mark the given task as OOM victim
750 * @tsk: task to mark
751 *
752 * Has to be called with oom_lock held and never after
753 * oom has been disabled already.
754 *
755 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
756 * under task_lock or operate on the current).
757 */
758 static void mark_oom_victim(struct task_struct *tsk)
759 {
760 struct mm_struct *mm = tsk->mm;
761
762 WARN_ON(oom_killer_disabled);
763 /* OOM killer might race with memcg OOM */
764 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
765 return;
766
767 /* oom_mm is bound to the signal struct life time. */
768 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
769 mmgrab(tsk->signal->oom_mm);
770
771 /*
772 * Make sure that the task is woken up from uninterruptible sleep
773 * if it is frozen because OOM killer wouldn't be able to free
774 * any memory and livelock. freezing_slow_path will tell the freezer
775 * that TIF_MEMDIE tasks should be ignored.
776 */
777 __thaw_task(tsk);
778 atomic_inc(&oom_victims);
779 trace_mark_victim(tsk->pid);
780 }
781
782 /**
783 * exit_oom_victim - note the exit of an OOM victim
784 */
exit_oom_victim(void)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 */
oom_killer_enable(void)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 */
oom_killer_disable(signed long timeout)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
__task_will_free_mem(struct task_struct * task)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 */
task_will_free_mem(struct task_struct * task)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
__oom_kill_process(struct task_struct * victim,const char * message)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 #undef K
998
999 /*
1000 * Kill provided task unless it's secured by setting
1001 * oom_score_adj to OOM_SCORE_ADJ_MIN.
1002 */
oom_kill_memcg_member(struct task_struct * task,void * message)1003 static int oom_kill_memcg_member(struct task_struct *task, void *message)
1004 {
1005 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1006 !is_global_init(task)) {
1007 get_task_struct(task);
1008 __oom_kill_process(task, message);
1009 }
1010 return 0;
1011 }
1012
oom_kill_process(struct oom_control * oc,const char * message)1013 static void oom_kill_process(struct oom_control *oc, const char *message)
1014 {
1015 struct task_struct *victim = oc->chosen;
1016 struct mem_cgroup *oom_group;
1017 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1018 DEFAULT_RATELIMIT_BURST);
1019
1020 /*
1021 * If the task is already exiting, don't alarm the sysadmin or kill
1022 * its children or threads, just give it access to memory reserves
1023 * so it can die quickly
1024 */
1025 task_lock(victim);
1026 if (task_will_free_mem(victim)) {
1027 mark_oom_victim(victim);
1028 queue_oom_reaper(victim);
1029 task_unlock(victim);
1030 put_task_struct(victim);
1031 return;
1032 }
1033 task_unlock(victim);
1034
1035 if (__ratelimit(&oom_rs))
1036 dump_header(oc, victim);
1037
1038 /*
1039 * Do we need to kill the entire memory cgroup?
1040 * Or even one of the ancestor memory cgroups?
1041 * Check this out before killing the victim task.
1042 */
1043 oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1044
1045 __oom_kill_process(victim, message);
1046
1047 /*
1048 * If necessary, kill all tasks in the selected memory cgroup.
1049 */
1050 if (oom_group) {
1051 memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1052 mem_cgroup_print_oom_group(oom_group);
1053 mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1054 (void *)message);
1055 mem_cgroup_put(oom_group);
1056 }
1057 }
1058
1059 /*
1060 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1061 */
check_panic_on_oom(struct oom_control * oc)1062 static void check_panic_on_oom(struct oom_control *oc)
1063 {
1064 if (likely(!sysctl_panic_on_oom))
1065 return;
1066 if (sysctl_panic_on_oom != 2) {
1067 /*
1068 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1069 * does not panic for cpuset, mempolicy, or memcg allocation
1070 * failures.
1071 */
1072 if (oc->constraint != CONSTRAINT_NONE)
1073 return;
1074 }
1075 /* Do not panic for oom kills triggered by sysrq */
1076 if (is_sysrq_oom(oc))
1077 return;
1078 dump_header(oc, NULL);
1079 panic("Out of memory: %s panic_on_oom is enabled\n",
1080 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1081 }
1082
1083 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1084
register_oom_notifier(struct notifier_block * nb)1085 int register_oom_notifier(struct notifier_block *nb)
1086 {
1087 return blocking_notifier_chain_register(&oom_notify_list, nb);
1088 }
1089 EXPORT_SYMBOL_GPL(register_oom_notifier);
1090
unregister_oom_notifier(struct notifier_block * nb)1091 int unregister_oom_notifier(struct notifier_block *nb)
1092 {
1093 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1094 }
1095 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1096
1097 /**
1098 * out_of_memory - kill the "best" process when we run out of memory
1099 * @oc: pointer to struct oom_control
1100 *
1101 * If we run out of memory, we have the choice between either
1102 * killing a random task (bad), letting the system crash (worse)
1103 * OR try to be smart about which process to kill. Note that we
1104 * don't have to be perfect here, we just have to be good.
1105 */
out_of_memory(struct oom_control * oc)1106 bool out_of_memory(struct oom_control *oc)
1107 {
1108 unsigned long freed = 0;
1109
1110 if (oom_killer_disabled)
1111 return false;
1112
1113 if (!is_memcg_oom(oc)) {
1114 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1115 if (freed > 0 && !is_sysrq_oom(oc))
1116 /* Got some memory back in the last second. */
1117 return true;
1118 }
1119
1120 /*
1121 * If current has a pending SIGKILL or is exiting, then automatically
1122 * select it. The goal is to allow it to allocate so that it may
1123 * quickly exit and free its memory.
1124 */
1125 if (task_will_free_mem(current)) {
1126 mark_oom_victim(current);
1127 queue_oom_reaper(current);
1128 return true;
1129 }
1130
1131 /*
1132 * The OOM killer does not compensate for IO-less reclaim.
1133 * pagefault_out_of_memory lost its gfp context so we have to
1134 * make sure exclude 0 mask - all other users should have at least
1135 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1136 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1137 */
1138 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1139 return true;
1140
1141 /*
1142 * Check if there were limitations on the allocation (only relevant for
1143 * NUMA and memcg) that may require different handling.
1144 */
1145 oc->constraint = constrained_alloc(oc);
1146 if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1147 oc->nodemask = NULL;
1148 check_panic_on_oom(oc);
1149
1150 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1151 current->mm && !oom_unkillable_task(current) &&
1152 oom_cpuset_eligible(current, oc) &&
1153 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1154 get_task_struct(current);
1155 oc->chosen = current;
1156 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1157 return true;
1158 }
1159
1160 select_bad_process(oc);
1161 /* Found nothing?!?! */
1162 if (!oc->chosen) {
1163 dump_header(oc, NULL);
1164 pr_warn("Out of memory and no killable processes...\n");
1165 /*
1166 * If we got here due to an actual allocation at the
1167 * system level, we cannot survive this and will enter
1168 * an endless loop in the allocator. Bail out now.
1169 */
1170 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1171 panic("System is deadlocked on memory\n");
1172 }
1173 if (oc->chosen && oc->chosen != (void *)-1UL)
1174 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1175 "Memory cgroup out of memory");
1176 return !!oc->chosen;
1177 }
1178
1179 /*
1180 * The pagefault handler calls here because some allocation has failed. We have
1181 * to take care of the memcg OOM here because this is the only safe context without
1182 * any locks held but let the oom killer triggered from the allocation context care
1183 * about the global OOM.
1184 */
pagefault_out_of_memory(void)1185 void pagefault_out_of_memory(void)
1186 {
1187 static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1188 DEFAULT_RATELIMIT_BURST);
1189
1190 if (mem_cgroup_oom_synchronize(true))
1191 return;
1192
1193 if (fatal_signal_pending(current))
1194 return;
1195
1196 if (__ratelimit(&pfoom_rs))
1197 pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1198 }
1199
SYSCALL_DEFINE2(process_mrelease,int,pidfd,unsigned int,flags)1200 SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1201 {
1202 #ifdef CONFIG_MMU
1203 struct mm_struct *mm = NULL;
1204 struct task_struct *task;
1205 struct task_struct *p;
1206 unsigned int f_flags;
1207 bool reap = false;
1208 long ret = 0;
1209
1210 if (flags)
1211 return -EINVAL;
1212
1213 task = pidfd_get_task(pidfd, &f_flags);
1214 if (IS_ERR(task))
1215 return PTR_ERR(task);
1216
1217 /*
1218 * Make sure to choose a thread which still has a reference to mm
1219 * during the group exit
1220 */
1221 p = find_lock_task_mm(task);
1222 if (!p) {
1223 ret = -ESRCH;
1224 goto put_task;
1225 }
1226
1227 mm = p->mm;
1228 mmgrab(mm);
1229
1230 if (task_will_free_mem(p))
1231 reap = true;
1232 else {
1233 /* Error only if the work has not been done already */
1234 if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1235 ret = -EINVAL;
1236 }
1237 task_unlock(p);
1238
1239 if (!reap)
1240 goto drop_mm;
1241
1242 if (mmap_read_lock_killable(mm)) {
1243 ret = -EINTR;
1244 goto drop_mm;
1245 }
1246 /*
1247 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1248 * possible change in exit_mmap is seen
1249 */
1250 if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1251 ret = -EAGAIN;
1252 mmap_read_unlock(mm);
1253
1254 drop_mm:
1255 mmdrop(mm);
1256 put_task:
1257 put_task_struct(task);
1258 return ret;
1259 #else
1260 return -ENOSYS;
1261 #endif /* CONFIG_MMU */
1262 }
1263