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
55static int sysctl_panic_on_oom;
56static int sysctl_oom_kill_allocating_task;
57static 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 */
67DEFINE_MUTEX(oom_lock);
68/* Serializes oom_score_adj and oom_score_adj_min updates */
69DEFINE_MUTEX(oom_adj_mutex);
70
71static 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 */
89static 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, tsk2: tsk);
112 }
113 if (ret)
114 break;
115 }
116 rcu_read_unlock();
117
118 return ret;
119}
120#else
121static 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 */
133struct 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(p: t);
141 if (likely(t->mm))
142 goto found;
143 task_unlock(p: t);
144 }
145 t = NULL;
146found:
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 */
156static 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. */
162static bool oom_unkillable_task(struct task_struct *p)
163{
164 if (is_global_init(tsk: 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*/
177static bool should_dump_unreclaim_slab(void)
178{
179 unsigned long nr_lru;
180
181 nr_lru = global_node_page_state(item: NR_ACTIVE_ANON) +
182 global_node_page_state(item: NR_INACTIVE_ANON) +
183 global_node_page_state(item: NR_ACTIVE_FILE) +
184 global_node_page_state(item: NR_INACTIVE_FILE) +
185 global_node_page_state(item: NR_ISOLATED_ANON) +
186 global_node_page_state(item: NR_ISOLATED_FILE) +
187 global_node_page_state(item: NR_UNEVICTABLE);
188
189 return (global_node_page_state_pages(item: 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 */
201long 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(tsk: 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(mm: p->mm) + get_mm_counter(mm: p->mm, member: MM_SWAPENTS) +
231 mm_pgtables_bytes(mm: 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
241static 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 */
251static 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(flags: 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(memcg: 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(z: zone, gfp_mask: 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
308static 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(p: task))
314 goto next;
315
316 /* p may not have freeable memory in nodemask */
317 if (!is_memcg_oom(oc) && !oom_cpuset_eligible(start: 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(tsk: 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(p: task)) {
337 points = LONG_MAX;
338 goto select;
339 }
340
341 points = oom_badness(p: task, totalpages: oc->totalpages);
342 if (points == LONG_MIN || points < oc->chosen_points)
343 goto next;
344
345select:
346 if (oc->chosen)
347 put_task_struct(t: oc->chosen);
348 get_task_struct(t: task);
349 oc->chosen = task;
350 oc->chosen_points = points;
351next:
352 return 0;
353abort:
354 if (oc->chosen)
355 put_task_struct(t: 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 */
364static 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(memcg: oc->memcg, oom_evaluate_task, arg: oc);
370 else {
371 struct task_struct *p;
372
373 rcu_read_lock();
374 for_each_process(p)
375 if (oom_evaluate_task(task: p, arg: oc))
376 break;
377 rcu_read_unlock();
378 }
379}
380
381static 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(start: 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(p: 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 */
423static 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(memcg: oc->memcg, dump_task, arg: oc);
430 else {
431 struct task_struct *p;
432
433 rcu_read_lock();
434 for_each_process(p)
435 dump_task(p, arg: oc);
436 rcu_read_unlock();
437 }
438}
439
440static void dump_oom_victim(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(memcg: oc->memcg, p: 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
452static void dump_header(struct oom_control *oc)
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(memcg: oc->memcg);
463 else {
464 __show_mem(SHOW_MEM_FILTER_NODES, nodemask: oc->nodemask, max_zone_idx: gfp_zone(flags: 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}
471
472/*
473 * Number of OOM victims in flight
474 */
475static atomic_t oom_victims = ATOMIC_INIT(0);
476static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
477
478static bool oom_killer_disabled __read_mostly;
479
480/*
481 * task->mm can be NULL if the task is the exited group leader. So to
482 * determine whether the task is using a particular mm, we examine all the
483 * task's threads: if one of those is using this mm then this task was also
484 * using it.
485 */
486bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
487{
488 struct task_struct *t;
489
490 for_each_thread(p, t) {
491 struct mm_struct *t_mm = READ_ONCE(t->mm);
492 if (t_mm)
493 return t_mm == mm;
494 }
495 return false;
496}
497
498#ifdef CONFIG_MMU
499/*
500 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
501 * victim (if that is possible) to help the OOM killer to move on.
502 */
503static struct task_struct *oom_reaper_th;
504static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
505static struct task_struct *oom_reaper_list;
506static DEFINE_SPINLOCK(oom_reaper_lock);
507
508static bool __oom_reap_task_mm(struct mm_struct *mm)
509{
510 struct vm_area_struct *vma;
511 bool ret = true;
512 VMA_ITERATOR(vmi, mm, 0);
513
514 /*
515 * Tell all users of get_user/copy_from_user etc... that the content
516 * is no longer stable. No barriers really needed because unmapping
517 * should imply barriers already and the reader would hit a page fault
518 * if it stumbled over a reaped memory.
519 */
520 set_bit(MMF_UNSTABLE, addr: &mm->flags);
521
522 for_each_vma(vmi, vma) {
523 if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
524 continue;
525
526 /*
527 * Only anonymous pages have a good chance to be dropped
528 * without additional steps which we cannot afford as we
529 * are OOM already.
530 *
531 * We do not even care about fs backed pages because all
532 * which are reclaimable have already been reclaimed and
533 * we do not want to block exit_mmap by keeping mm ref
534 * count elevated without a good reason.
535 */
536 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
537 struct mmu_notifier_range range;
538 struct mmu_gather tlb;
539
540 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_UNMAP, flags: 0,
541 mm, start: vma->vm_start,
542 end: vma->vm_end);
543 tlb_gather_mmu(tlb: &tlb, mm);
544 if (mmu_notifier_invalidate_range_start_nonblock(range: &range)) {
545 tlb_finish_mmu(tlb: &tlb);
546 ret = false;
547 continue;
548 }
549 unmap_page_range(tlb: &tlb, vma, addr: range.start, end: range.end, NULL);
550 mmu_notifier_invalidate_range_end(range: &range);
551 tlb_finish_mmu(tlb: &tlb);
552 }
553 }
554
555 return ret;
556}
557
558/*
559 * Reaps the address space of the give task.
560 *
561 * Returns true on success and false if none or part of the address space
562 * has been reclaimed and the caller should retry later.
563 */
564static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
565{
566 bool ret = true;
567
568 if (!mmap_read_trylock(mm)) {
569 trace_skip_task_reaping(pid: tsk->pid);
570 return false;
571 }
572
573 /*
574 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
575 * work on the mm anymore. The check for MMF_OOM_SKIP must run
576 * under mmap_lock for reading because it serializes against the
577 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
578 */
579 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
580 trace_skip_task_reaping(pid: tsk->pid);
581 goto out_unlock;
582 }
583
584 trace_start_task_reaping(pid: tsk->pid);
585
586 /* failed to reap part of the address space. Try again later */
587 ret = __oom_reap_task_mm(mm);
588 if (!ret)
589 goto out_finish;
590
591 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
592 task_pid_nr(tsk), tsk->comm,
593 K(get_mm_counter(mm, MM_ANONPAGES)),
594 K(get_mm_counter(mm, MM_FILEPAGES)),
595 K(get_mm_counter(mm, MM_SHMEMPAGES)));
596out_finish:
597 trace_finish_task_reaping(pid: tsk->pid);
598out_unlock:
599 mmap_read_unlock(mm);
600
601 return ret;
602}
603
604#define MAX_OOM_REAP_RETRIES 10
605static void oom_reap_task(struct task_struct *tsk)
606{
607 int attempts = 0;
608 struct mm_struct *mm = tsk->signal->oom_mm;
609
610 /* Retry the mmap_read_trylock(mm) a few times */
611 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
612 schedule_timeout_idle(HZ/10);
613
614 if (attempts <= MAX_OOM_REAP_RETRIES ||
615 test_bit(MMF_OOM_SKIP, &mm->flags))
616 goto done;
617
618 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
619 task_pid_nr(tsk), tsk->comm);
620 sched_show_task(p: tsk);
621 debug_show_all_locks();
622
623done:
624 tsk->oom_reaper_list = NULL;
625
626 /*
627 * Hide this mm from OOM killer because it has been either reaped or
628 * somebody can't call mmap_write_unlock(mm).
629 */
630 set_bit(MMF_OOM_SKIP, addr: &mm->flags);
631
632 /* Drop a reference taken by queue_oom_reaper */
633 put_task_struct(t: tsk);
634}
635
636static int oom_reaper(void *unused)
637{
638 set_freezable();
639
640 while (true) {
641 struct task_struct *tsk = NULL;
642
643 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
644 spin_lock_irq(lock: &oom_reaper_lock);
645 if (oom_reaper_list != NULL) {
646 tsk = oom_reaper_list;
647 oom_reaper_list = tsk->oom_reaper_list;
648 }
649 spin_unlock_irq(lock: &oom_reaper_lock);
650
651 if (tsk)
652 oom_reap_task(tsk);
653 }
654
655 return 0;
656}
657
658static void wake_oom_reaper(struct timer_list *timer)
659{
660 struct task_struct *tsk = container_of(timer, struct task_struct,
661 oom_reaper_timer);
662 struct mm_struct *mm = tsk->signal->oom_mm;
663 unsigned long flags;
664
665 /* The victim managed to terminate on its own - see exit_mmap */
666 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
667 put_task_struct(t: tsk);
668 return;
669 }
670
671 spin_lock_irqsave(&oom_reaper_lock, flags);
672 tsk->oom_reaper_list = oom_reaper_list;
673 oom_reaper_list = tsk;
674 spin_unlock_irqrestore(lock: &oom_reaper_lock, flags);
675 trace_wake_reaper(pid: tsk->pid);
676 wake_up(&oom_reaper_wait);
677}
678
679/*
680 * Give the OOM victim time to exit naturally before invoking the oom_reaping.
681 * The timers timeout is arbitrary... the longer it is, the longer the worst
682 * case scenario for the OOM can take. If it is too small, the oom_reaper can
683 * get in the way and release resources needed by the process exit path.
684 * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
685 * before the exit path is able to wake the futex waiters.
686 */
687#define OOM_REAPER_DELAY (2*HZ)
688static void queue_oom_reaper(struct task_struct *tsk)
689{
690 /* mm is already queued? */
691 if (test_and_set_bit(MMF_OOM_REAP_QUEUED, addr: &tsk->signal->oom_mm->flags))
692 return;
693
694 get_task_struct(t: tsk);
695 timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
696 tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
697 add_timer(timer: &tsk->oom_reaper_timer);
698}
699
700#ifdef CONFIG_SYSCTL
701static struct ctl_table vm_oom_kill_table[] = {
702 {
703 .procname = "panic_on_oom",
704 .data = &sysctl_panic_on_oom,
705 .maxlen = sizeof(sysctl_panic_on_oom),
706 .mode = 0644,
707 .proc_handler = proc_dointvec_minmax,
708 .extra1 = SYSCTL_ZERO,
709 .extra2 = SYSCTL_TWO,
710 },
711 {
712 .procname = "oom_kill_allocating_task",
713 .data = &sysctl_oom_kill_allocating_task,
714 .maxlen = sizeof(sysctl_oom_kill_allocating_task),
715 .mode = 0644,
716 .proc_handler = proc_dointvec,
717 },
718 {
719 .procname = "oom_dump_tasks",
720 .data = &sysctl_oom_dump_tasks,
721 .maxlen = sizeof(sysctl_oom_dump_tasks),
722 .mode = 0644,
723 .proc_handler = proc_dointvec,
724 },
725 {}
726};
727#endif
728
729static int __init oom_init(void)
730{
731 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
732#ifdef CONFIG_SYSCTL
733 register_sysctl_init("vm", vm_oom_kill_table);
734#endif
735 return 0;
736}
737subsys_initcall(oom_init)
738#else
739static inline void queue_oom_reaper(struct task_struct *tsk)
740{
741}
742#endif /* CONFIG_MMU */
743
744/**
745 * mark_oom_victim - mark the given task as OOM victim
746 * @tsk: task to mark
747 *
748 * Has to be called with oom_lock held and never after
749 * oom has been disabled already.
750 *
751 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
752 * under task_lock or operate on the current).
753 */
754static void mark_oom_victim(struct task_struct *tsk)
755{
756 struct mm_struct *mm = tsk->mm;
757
758 WARN_ON(oom_killer_disabled);
759 /* OOM killer might race with memcg OOM */
760 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
761 return;
762
763 /* oom_mm is bound to the signal struct life time. */
764 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
765 mmgrab(mm: tsk->signal->oom_mm);
766
767 /*
768 * Make sure that the task is woken up from uninterruptible sleep
769 * if it is frozen because OOM killer wouldn't be able to free
770 * any memory and livelock. freezing_slow_path will tell the freezer
771 * that TIF_MEMDIE tasks should be ignored.
772 */
773 __thaw_task(t: tsk);
774 atomic_inc(v: &oom_victims);
775 trace_mark_victim(pid: tsk->pid);
776}
777
778/**
779 * exit_oom_victim - note the exit of an OOM victim
780 */
781void exit_oom_victim(void)
782{
783 clear_thread_flag(TIF_MEMDIE);
784
785 if (!atomic_dec_return(v: &oom_victims))
786 wake_up_all(&oom_victims_wait);
787}
788
789/**
790 * oom_killer_enable - enable OOM killer
791 */
792void oom_killer_enable(void)
793{
794 oom_killer_disabled = false;
795 pr_info("OOM killer enabled.\n");
796}
797
798/**
799 * oom_killer_disable - disable OOM killer
800 * @timeout: maximum timeout to wait for oom victims in jiffies
801 *
802 * Forces all page allocations to fail rather than trigger OOM killer.
803 * Will block and wait until all OOM victims are killed or the given
804 * timeout expires.
805 *
806 * The function cannot be called when there are runnable user tasks because
807 * the userspace would see unexpected allocation failures as a result. Any
808 * new usage of this function should be consulted with MM people.
809 *
810 * Returns true if successful and false if the OOM killer cannot be
811 * disabled.
812 */
813bool oom_killer_disable(signed long timeout)
814{
815 signed long ret;
816
817 /*
818 * Make sure to not race with an ongoing OOM killer. Check that the
819 * current is not killed (possibly due to sharing the victim's memory).
820 */
821 if (mutex_lock_killable(&oom_lock))
822 return false;
823 oom_killer_disabled = true;
824 mutex_unlock(lock: &oom_lock);
825
826 ret = wait_event_interruptible_timeout(oom_victims_wait,
827 !atomic_read(&oom_victims), timeout);
828 if (ret <= 0) {
829 oom_killer_enable();
830 return false;
831 }
832 pr_info("OOM killer disabled.\n");
833
834 return true;
835}
836
837static inline bool __task_will_free_mem(struct task_struct *task)
838{
839 struct signal_struct *sig = task->signal;
840
841 /*
842 * A coredumping process may sleep for an extended period in
843 * coredump_task_exit(), so the oom killer cannot assume that
844 * the process will promptly exit and release memory.
845 */
846 if (sig->core_state)
847 return false;
848
849 if (sig->flags & SIGNAL_GROUP_EXIT)
850 return true;
851
852 if (thread_group_empty(p: task) && (task->flags & PF_EXITING))
853 return true;
854
855 return false;
856}
857
858/*
859 * Checks whether the given task is dying or exiting and likely to
860 * release its address space. This means that all threads and processes
861 * sharing the same mm have to be killed or exiting.
862 * Caller has to make sure that task->mm is stable (hold task_lock or
863 * it operates on the current).
864 */
865static bool task_will_free_mem(struct task_struct *task)
866{
867 struct mm_struct *mm = task->mm;
868 struct task_struct *p;
869 bool ret = true;
870
871 /*
872 * Skip tasks without mm because it might have passed its exit_mm and
873 * exit_oom_victim. oom_reaper could have rescued that but do not rely
874 * on that for now. We can consider find_lock_task_mm in future.
875 */
876 if (!mm)
877 return false;
878
879 if (!__task_will_free_mem(task))
880 return false;
881
882 /*
883 * This task has already been drained by the oom reaper so there are
884 * only small chances it will free some more
885 */
886 if (test_bit(MMF_OOM_SKIP, &mm->flags))
887 return false;
888
889 if (atomic_read(v: &mm->mm_users) <= 1)
890 return true;
891
892 /*
893 * Make sure that all tasks which share the mm with the given tasks
894 * are dying as well to make sure that a) nobody pins its mm and
895 * b) the task is also reapable by the oom reaper.
896 */
897 rcu_read_lock();
898 for_each_process(p) {
899 if (!process_shares_mm(p, mm))
900 continue;
901 if (same_thread_group(p1: task, p2: p))
902 continue;
903 ret = __task_will_free_mem(task: p);
904 if (!ret)
905 break;
906 }
907 rcu_read_unlock();
908
909 return ret;
910}
911
912static void __oom_kill_process(struct task_struct *victim, const char *message)
913{
914 struct task_struct *p;
915 struct mm_struct *mm;
916 bool can_oom_reap = true;
917
918 p = find_lock_task_mm(p: victim);
919 if (!p) {
920 pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
921 message, task_pid_nr(victim), victim->comm);
922 put_task_struct(t: victim);
923 return;
924 } else if (victim != p) {
925 get_task_struct(t: p);
926 put_task_struct(t: victim);
927 victim = p;
928 }
929
930 /* Get a reference to safely compare mm after task_unlock(victim) */
931 mm = victim->mm;
932 mmgrab(mm);
933
934 /* Raise event before sending signal: task reaper must see this */
935 count_vm_event(item: OOM_KILL);
936 memcg_memory_event_mm(mm, event: MEMCG_OOM_KILL);
937
938 /*
939 * We should send SIGKILL before granting access to memory reserves
940 * in order to prevent the OOM victim from depleting the memory
941 * reserves from the user space under its control.
942 */
943 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p: victim, type: PIDTYPE_TGID);
944 mark_oom_victim(tsk: victim);
945 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",
946 message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
947 K(get_mm_counter(mm, MM_ANONPAGES)),
948 K(get_mm_counter(mm, MM_FILEPAGES)),
949 K(get_mm_counter(mm, MM_SHMEMPAGES)),
950 from_kuid(&init_user_ns, task_uid(victim)),
951 mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
952 task_unlock(p: victim);
953
954 /*
955 * Kill all user processes sharing victim->mm in other thread groups, if
956 * any. They don't get access to memory reserves, though, to avoid
957 * depletion of all memory. This prevents mm->mmap_lock livelock when an
958 * oom killed thread cannot exit because it requires the semaphore and
959 * its contended by another thread trying to allocate memory itself.
960 * That thread will now get access to memory reserves since it has a
961 * pending fatal signal.
962 */
963 rcu_read_lock();
964 for_each_process(p) {
965 if (!process_shares_mm(p, mm))
966 continue;
967 if (same_thread_group(p1: p, p2: victim))
968 continue;
969 if (is_global_init(tsk: p)) {
970 can_oom_reap = false;
971 set_bit(MMF_OOM_SKIP, addr: &mm->flags);
972 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
973 task_pid_nr(victim), victim->comm,
974 task_pid_nr(p), p->comm);
975 continue;
976 }
977 /*
978 * No kthread_use_mm() user needs to read from the userspace so
979 * we are ok to reap it.
980 */
981 if (unlikely(p->flags & PF_KTHREAD))
982 continue;
983 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, type: PIDTYPE_TGID);
984 }
985 rcu_read_unlock();
986
987 if (can_oom_reap)
988 queue_oom_reaper(tsk: victim);
989
990 mmdrop(mm);
991 put_task_struct(t: victim);
992}
993
994/*
995 * Kill provided task unless it's secured by setting
996 * oom_score_adj to OOM_SCORE_ADJ_MIN.
997 */
998static int oom_kill_memcg_member(struct task_struct *task, void *message)
999{
1000 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1001 !is_global_init(tsk: task)) {
1002 get_task_struct(t: task);
1003 __oom_kill_process(victim: task, message);
1004 }
1005 return 0;
1006}
1007
1008static void oom_kill_process(struct oom_control *oc, const char *message)
1009{
1010 struct task_struct *victim = oc->chosen;
1011 struct mem_cgroup *oom_group;
1012 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1013 DEFAULT_RATELIMIT_BURST);
1014
1015 /*
1016 * If the task is already exiting, don't alarm the sysadmin or kill
1017 * its children or threads, just give it access to memory reserves
1018 * so it can die quickly
1019 */
1020 task_lock(p: victim);
1021 if (task_will_free_mem(task: victim)) {
1022 mark_oom_victim(tsk: victim);
1023 queue_oom_reaper(tsk: victim);
1024 task_unlock(p: victim);
1025 put_task_struct(t: victim);
1026 return;
1027 }
1028 task_unlock(p: victim);
1029
1030 if (__ratelimit(&oom_rs)) {
1031 dump_header(oc);
1032 dump_oom_victim(oc, victim);
1033 }
1034
1035 /*
1036 * Do we need to kill the entire memory cgroup?
1037 * Or even one of the ancestor memory cgroups?
1038 * Check this out before killing the victim task.
1039 */
1040 oom_group = mem_cgroup_get_oom_group(victim, oom_domain: oc->memcg);
1041
1042 __oom_kill_process(victim, message);
1043
1044 /*
1045 * If necessary, kill all tasks in the selected memory cgroup.
1046 */
1047 if (oom_group) {
1048 memcg_memory_event(memcg: oom_group, event: MEMCG_OOM_GROUP_KILL);
1049 mem_cgroup_print_oom_group(memcg: oom_group);
1050 mem_cgroup_scan_tasks(memcg: oom_group, oom_kill_memcg_member,
1051 arg: (void *)message);
1052 mem_cgroup_put(memcg: oom_group);
1053 }
1054}
1055
1056/*
1057 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1058 */
1059static void check_panic_on_oom(struct oom_control *oc)
1060{
1061 if (likely(!sysctl_panic_on_oom))
1062 return;
1063 if (sysctl_panic_on_oom != 2) {
1064 /*
1065 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1066 * does not panic for cpuset, mempolicy, or memcg allocation
1067 * failures.
1068 */
1069 if (oc->constraint != CONSTRAINT_NONE)
1070 return;
1071 }
1072 /* Do not panic for oom kills triggered by sysrq */
1073 if (is_sysrq_oom(oc))
1074 return;
1075 dump_header(oc);
1076 panic(fmt: "Out of memory: %s panic_on_oom is enabled\n",
1077 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1078}
1079
1080static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1081
1082int register_oom_notifier(struct notifier_block *nb)
1083{
1084 return blocking_notifier_chain_register(nh: &oom_notify_list, nb);
1085}
1086EXPORT_SYMBOL_GPL(register_oom_notifier);
1087
1088int unregister_oom_notifier(struct notifier_block *nb)
1089{
1090 return blocking_notifier_chain_unregister(nh: &oom_notify_list, nb);
1091}
1092EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1093
1094/**
1095 * out_of_memory - kill the "best" process when we run out of memory
1096 * @oc: pointer to struct oom_control
1097 *
1098 * If we run out of memory, we have the choice between either
1099 * killing a random task (bad), letting the system crash (worse)
1100 * OR try to be smart about which process to kill. Note that we
1101 * don't have to be perfect here, we just have to be good.
1102 */
1103bool out_of_memory(struct oom_control *oc)
1104{
1105 unsigned long freed = 0;
1106
1107 if (oom_killer_disabled)
1108 return false;
1109
1110 if (!is_memcg_oom(oc)) {
1111 blocking_notifier_call_chain(nh: &oom_notify_list, val: 0, v: &freed);
1112 if (freed > 0 && !is_sysrq_oom(oc))
1113 /* Got some memory back in the last second. */
1114 return true;
1115 }
1116
1117 /*
1118 * If current has a pending SIGKILL or is exiting, then automatically
1119 * select it. The goal is to allow it to allocate so that it may
1120 * quickly exit and free its memory.
1121 */
1122 if (task_will_free_mem(current)) {
1123 mark_oom_victim(current);
1124 queue_oom_reaper(current);
1125 return true;
1126 }
1127
1128 /*
1129 * The OOM killer does not compensate for IO-less reclaim.
1130 * But mem_cgroup_oom() has to invoke the OOM killer even
1131 * if it is a GFP_NOFS allocation.
1132 */
1133 if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1134 return true;
1135
1136 /*
1137 * Check if there were limitations on the allocation (only relevant for
1138 * NUMA and memcg) that may require different handling.
1139 */
1140 oc->constraint = constrained_alloc(oc);
1141 if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1142 oc->nodemask = NULL;
1143 check_panic_on_oom(oc);
1144
1145 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1146 current->mm && !oom_unkillable_task(current) &&
1147 oom_cpuset_eligible(current, oc) &&
1148 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1149 get_task_struct(current);
1150 oc->chosen = current;
1151 oom_kill_process(oc, message: "Out of memory (oom_kill_allocating_task)");
1152 return true;
1153 }
1154
1155 select_bad_process(oc);
1156 /* Found nothing?!?! */
1157 if (!oc->chosen) {
1158 dump_header(oc);
1159 pr_warn("Out of memory and no killable processes...\n");
1160 /*
1161 * If we got here due to an actual allocation at the
1162 * system level, we cannot survive this and will enter
1163 * an endless loop in the allocator. Bail out now.
1164 */
1165 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1166 panic(fmt: "System is deadlocked on memory\n");
1167 }
1168 if (oc->chosen && oc->chosen != (void *)-1UL)
1169 oom_kill_process(oc, message: !is_memcg_oom(oc) ? "Out of memory" :
1170 "Memory cgroup out of memory");
1171 return !!oc->chosen;
1172}
1173
1174/*
1175 * The pagefault handler calls here because some allocation has failed. We have
1176 * to take care of the memcg OOM here because this is the only safe context without
1177 * any locks held but let the oom killer triggered from the allocation context care
1178 * about the global OOM.
1179 */
1180void pagefault_out_of_memory(void)
1181{
1182 static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1183 DEFAULT_RATELIMIT_BURST);
1184
1185 if (mem_cgroup_oom_synchronize(wait: true))
1186 return;
1187
1188 if (fatal_signal_pending(current))
1189 return;
1190
1191 if (__ratelimit(&pfoom_rs))
1192 pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1193}
1194
1195SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1196{
1197#ifdef CONFIG_MMU
1198 struct mm_struct *mm = NULL;
1199 struct task_struct *task;
1200 struct task_struct *p;
1201 unsigned int f_flags;
1202 bool reap = false;
1203 long ret = 0;
1204
1205 if (flags)
1206 return -EINVAL;
1207
1208 task = pidfd_get_task(pidfd, flags: &f_flags);
1209 if (IS_ERR(ptr: task))
1210 return PTR_ERR(ptr: task);
1211
1212 /*
1213 * Make sure to choose a thread which still has a reference to mm
1214 * during the group exit
1215 */
1216 p = find_lock_task_mm(p: task);
1217 if (!p) {
1218 ret = -ESRCH;
1219 goto put_task;
1220 }
1221
1222 mm = p->mm;
1223 mmgrab(mm);
1224
1225 if (task_will_free_mem(task: p))
1226 reap = true;
1227 else {
1228 /* Error only if the work has not been done already */
1229 if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1230 ret = -EINVAL;
1231 }
1232 task_unlock(p);
1233
1234 if (!reap)
1235 goto drop_mm;
1236
1237 if (mmap_read_lock_killable(mm)) {
1238 ret = -EINTR;
1239 goto drop_mm;
1240 }
1241 /*
1242 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1243 * possible change in exit_mmap is seen
1244 */
1245 if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1246 ret = -EAGAIN;
1247 mmap_read_unlock(mm);
1248
1249drop_mm:
1250 mmdrop(mm);
1251put_task:
1252 put_task_struct(t: task);
1253 return ret;
1254#else
1255 return -ENOSYS;
1256#endif /* CONFIG_MMU */
1257}
1258

source code of linux/mm/oom_kill.c