Loading...
1/*
2 * linux/mm/oom_kill.c
3 *
4 * Copyright (C) 1998,2000 Rik van Riel
5 * Thanks go out to Claus Fischer for some serious inspiration and
6 * for goading me into coding this file...
7 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
9 *
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
13 *
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
18 */
19
20#include <linux/oom.h>
21#include <linux/mm.h>
22#include <linux/err.h>
23#include <linux/gfp.h>
24#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
28#include <linux/cpuset.h>
29#include <linux/module.h>
30#include <linux/notifier.h>
31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
33#include <linux/security.h>
34#include <linux/ptrace.h>
35
36int sysctl_panic_on_oom;
37int sysctl_oom_kill_allocating_task;
38int sysctl_oom_dump_tasks = 1;
39static DEFINE_SPINLOCK(zone_scan_lock);
40
41/**
42 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
43 * @new_val: new oom_score_adj value
44 *
45 * Sets the oom_score_adj value for current to @new_val with proper
46 * synchronization and returns the old value. Usually used to temporarily
47 * set a value, save the old value in the caller, and then reinstate it later.
48 */
49int test_set_oom_score_adj(int new_val)
50{
51 struct sighand_struct *sighand = current->sighand;
52 int old_val;
53
54 spin_lock_irq(&sighand->siglock);
55 old_val = current->signal->oom_score_adj;
56 if (new_val != old_val) {
57 if (new_val == OOM_SCORE_ADJ_MIN)
58 atomic_inc(¤t->mm->oom_disable_count);
59 else if (old_val == OOM_SCORE_ADJ_MIN)
60 atomic_dec(¤t->mm->oom_disable_count);
61 current->signal->oom_score_adj = new_val;
62 }
63 spin_unlock_irq(&sighand->siglock);
64
65 return old_val;
66}
67
68#ifdef CONFIG_NUMA
69/**
70 * has_intersects_mems_allowed() - check task eligiblity for kill
71 * @tsk: task struct of which task to consider
72 * @mask: nodemask passed to page allocator for mempolicy ooms
73 *
74 * Task eligibility is determined by whether or not a candidate task, @tsk,
75 * shares the same mempolicy nodes as current if it is bound by such a policy
76 * and whether or not it has the same set of allowed cpuset nodes.
77 */
78static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
80{
81 struct task_struct *start = tsk;
82
83 do {
84 if (mask) {
85 /*
86 * If this is a mempolicy constrained oom, tsk's
87 * cpuset is irrelevant. Only return true if its
88 * mempolicy intersects current, otherwise it may be
89 * needlessly killed.
90 */
91 if (mempolicy_nodemask_intersects(tsk, mask))
92 return true;
93 } else {
94 /*
95 * This is not a mempolicy constrained oom, so only
96 * check the mems of tsk's cpuset.
97 */
98 if (cpuset_mems_allowed_intersects(current, tsk))
99 return true;
100 }
101 } while_each_thread(start, tsk);
102
103 return false;
104}
105#else
106static bool has_intersects_mems_allowed(struct task_struct *tsk,
107 const nodemask_t *mask)
108{
109 return true;
110}
111#endif /* CONFIG_NUMA */
112
113/*
114 * The process p may have detached its own ->mm while exiting or through
115 * use_mm(), but one or more of its subthreads may still have a valid
116 * pointer. Return p, or any of its subthreads with a valid ->mm, with
117 * task_lock() held.
118 */
119struct task_struct *find_lock_task_mm(struct task_struct *p)
120{
121 struct task_struct *t = p;
122
123 do {
124 task_lock(t);
125 if (likely(t->mm))
126 return t;
127 task_unlock(t);
128 } while_each_thread(p, t);
129
130 return NULL;
131}
132
133/* return true if the task is not adequate as candidate victim task. */
134static bool oom_unkillable_task(struct task_struct *p,
135 const struct mem_cgroup *mem, const nodemask_t *nodemask)
136{
137 if (is_global_init(p))
138 return true;
139 if (p->flags & PF_KTHREAD)
140 return true;
141
142 /* When mem_cgroup_out_of_memory() and p is not member of the group */
143 if (mem && !task_in_mem_cgroup(p, mem))
144 return true;
145
146 /* p may not have freeable memory in nodemask */
147 if (!has_intersects_mems_allowed(p, nodemask))
148 return true;
149
150 return false;
151}
152
153/**
154 * oom_badness - heuristic function to determine which candidate task to kill
155 * @p: task struct of which task we should calculate
156 * @totalpages: total present RAM allowed for page allocation
157 *
158 * The heuristic for determining which task to kill is made to be as simple and
159 * predictable as possible. The goal is to return the highest value for the
160 * task consuming the most memory to avoid subsequent oom failures.
161 */
162unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
163 const nodemask_t *nodemask, unsigned long totalpages)
164{
165 int points;
166
167 if (oom_unkillable_task(p, mem, nodemask))
168 return 0;
169
170 p = find_lock_task_mm(p);
171 if (!p)
172 return 0;
173
174 /*
175 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
176 * so the entire heuristic doesn't need to be executed for something
177 * that cannot be killed.
178 */
179 if (atomic_read(&p->mm->oom_disable_count)) {
180 task_unlock(p);
181 return 0;
182 }
183
184 /*
185 * The memory controller may have a limit of 0 bytes, so avoid a divide
186 * by zero, if necessary.
187 */
188 if (!totalpages)
189 totalpages = 1;
190
191 /*
192 * The baseline for the badness score is the proportion of RAM that each
193 * task's rss, pagetable and swap space use.
194 */
195 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
196 points += get_mm_counter(p->mm, MM_SWAPENTS);
197
198 points *= 1000;
199 points /= totalpages;
200 task_unlock(p);
201
202 /*
203 * Root processes get 3% bonus, just like the __vm_enough_memory()
204 * implementation used by LSMs.
205 */
206 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
207 points -= 30;
208
209 /*
210 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
211 * either completely disable oom killing or always prefer a certain
212 * task.
213 */
214 points += p->signal->oom_score_adj;
215
216 /*
217 * Never return 0 for an eligible task that may be killed since it's
218 * possible that no single user task uses more than 0.1% of memory and
219 * no single admin tasks uses more than 3.0%.
220 */
221 if (points <= 0)
222 return 1;
223 return (points < 1000) ? points : 1000;
224}
225
226/*
227 * Determine the type of allocation constraint.
228 */
229#ifdef CONFIG_NUMA
230static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
231 gfp_t gfp_mask, nodemask_t *nodemask,
232 unsigned long *totalpages)
233{
234 struct zone *zone;
235 struct zoneref *z;
236 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
237 bool cpuset_limited = false;
238 int nid;
239
240 /* Default to all available memory */
241 *totalpages = totalram_pages + total_swap_pages;
242
243 if (!zonelist)
244 return CONSTRAINT_NONE;
245 /*
246 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
247 * to kill current.We have to random task kill in this case.
248 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
249 */
250 if (gfp_mask & __GFP_THISNODE)
251 return CONSTRAINT_NONE;
252
253 /*
254 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
255 * the page allocator means a mempolicy is in effect. Cpuset policy
256 * is enforced in get_page_from_freelist().
257 */
258 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
259 *totalpages = total_swap_pages;
260 for_each_node_mask(nid, *nodemask)
261 *totalpages += node_spanned_pages(nid);
262 return CONSTRAINT_MEMORY_POLICY;
263 }
264
265 /* Check this allocation failure is caused by cpuset's wall function */
266 for_each_zone_zonelist_nodemask(zone, z, zonelist,
267 high_zoneidx, nodemask)
268 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
269 cpuset_limited = true;
270
271 if (cpuset_limited) {
272 *totalpages = total_swap_pages;
273 for_each_node_mask(nid, cpuset_current_mems_allowed)
274 *totalpages += node_spanned_pages(nid);
275 return CONSTRAINT_CPUSET;
276 }
277 return CONSTRAINT_NONE;
278}
279#else
280static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281 gfp_t gfp_mask, nodemask_t *nodemask,
282 unsigned long *totalpages)
283{
284 *totalpages = totalram_pages + total_swap_pages;
285 return CONSTRAINT_NONE;
286}
287#endif
288
289/*
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. We expect the caller will lock the tasklist.
292 *
293 * (not docbooked, we don't want this one cluttering up the manual)
294 */
295static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, struct mem_cgroup *mem,
297 const nodemask_t *nodemask)
298{
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
301 *ppoints = 0;
302
303 do_each_thread(g, p) {
304 unsigned int points;
305
306 if (p->exit_state)
307 continue;
308 if (oom_unkillable_task(p, mem, nodemask))
309 continue;
310
311 /*
312 * This task already has access to memory reserves and is
313 * being killed. Don't allow any other task access to the
314 * memory reserve.
315 *
316 * Note: this may have a chance of deadlock if it gets
317 * blocked waiting for another task which itself is waiting
318 * for memory. Is there a better alternative?
319 */
320 if (test_tsk_thread_flag(p, TIF_MEMDIE))
321 return ERR_PTR(-1UL);
322 if (!p->mm)
323 continue;
324
325 if (p->flags & PF_EXITING) {
326 /*
327 * If p is the current task and is in the process of
328 * releasing memory, we allow the "kill" to set
329 * TIF_MEMDIE, which will allow it to gain access to
330 * memory reserves. Otherwise, it may stall forever.
331 *
332 * The loop isn't broken here, however, in case other
333 * threads are found to have already been oom killed.
334 */
335 if (p == current) {
336 chosen = p;
337 *ppoints = 1000;
338 } else {
339 /*
340 * If this task is not being ptraced on exit,
341 * then wait for it to finish before killing
342 * some other task unnecessarily.
343 */
344 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
345 return ERR_PTR(-1UL);
346 }
347 }
348
349 points = oom_badness(p, mem, nodemask, totalpages);
350 if (points > *ppoints) {
351 chosen = p;
352 *ppoints = points;
353 }
354 } while_each_thread(g, p);
355
356 return chosen;
357}
358
359/**
360 * dump_tasks - dump current memory state of all system tasks
361 * @mem: current's memory controller, if constrained
362 * @nodemask: nodemask passed to page allocator for mempolicy ooms
363 *
364 * Dumps the current memory state of all eligible tasks. Tasks not in the same
365 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
366 * are not shown.
367 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
368 * value, oom_score_adj value, and name.
369 *
370 * Call with tasklist_lock read-locked.
371 */
372static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
373{
374 struct task_struct *p;
375 struct task_struct *task;
376
377 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
378 for_each_process(p) {
379 if (oom_unkillable_task(p, mem, nodemask))
380 continue;
381
382 task = find_lock_task_mm(p);
383 if (!task) {
384 /*
385 * This is a kthread or all of p's threads have already
386 * detached their mm's. There's no need to report
387 * them; they can't be oom killed anyway.
388 */
389 continue;
390 }
391
392 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
393 task->pid, task_uid(task), task->tgid,
394 task->mm->total_vm, get_mm_rss(task->mm),
395 task_cpu(task), task->signal->oom_adj,
396 task->signal->oom_score_adj, task->comm);
397 task_unlock(task);
398 }
399}
400
401static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
402 struct mem_cgroup *mem, const nodemask_t *nodemask)
403{
404 task_lock(current);
405 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
406 "oom_adj=%d, oom_score_adj=%d\n",
407 current->comm, gfp_mask, order, current->signal->oom_adj,
408 current->signal->oom_score_adj);
409 cpuset_print_task_mems_allowed(current);
410 task_unlock(current);
411 dump_stack();
412 mem_cgroup_print_oom_info(mem, p);
413 show_mem(SHOW_MEM_FILTER_NODES);
414 if (sysctl_oom_dump_tasks)
415 dump_tasks(mem, nodemask);
416}
417
418#define K(x) ((x) << (PAGE_SHIFT-10))
419static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
420{
421 struct task_struct *q;
422 struct mm_struct *mm;
423
424 p = find_lock_task_mm(p);
425 if (!p)
426 return 1;
427
428 /* mm cannot be safely dereferenced after task_unlock(p) */
429 mm = p->mm;
430
431 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
432 task_pid_nr(p), p->comm, K(p->mm->total_vm),
433 K(get_mm_counter(p->mm, MM_ANONPAGES)),
434 K(get_mm_counter(p->mm, MM_FILEPAGES)));
435 task_unlock(p);
436
437 /*
438 * Kill all processes sharing p->mm in other thread groups, if any.
439 * They don't get access to memory reserves or a higher scheduler
440 * priority, though, to avoid depletion of all memory or task
441 * starvation. This prevents mm->mmap_sem livelock when an oom killed
442 * task cannot exit because it requires the semaphore and its contended
443 * by another thread trying to allocate memory itself. That thread will
444 * now get access to memory reserves since it has a pending fatal
445 * signal.
446 */
447 for_each_process(q)
448 if (q->mm == mm && !same_thread_group(q, p)) {
449 task_lock(q); /* Protect ->comm from prctl() */
450 pr_err("Kill process %d (%s) sharing same memory\n",
451 task_pid_nr(q), q->comm);
452 task_unlock(q);
453 force_sig(SIGKILL, q);
454 }
455
456 set_tsk_thread_flag(p, TIF_MEMDIE);
457 force_sig(SIGKILL, p);
458
459 return 0;
460}
461#undef K
462
463static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
464 unsigned int points, unsigned long totalpages,
465 struct mem_cgroup *mem, nodemask_t *nodemask,
466 const char *message)
467{
468 struct task_struct *victim = p;
469 struct task_struct *child;
470 struct task_struct *t = p;
471 unsigned int victim_points = 0;
472
473 if (printk_ratelimit())
474 dump_header(p, gfp_mask, order, mem, nodemask);
475
476 /*
477 * If the task is already exiting, don't alarm the sysadmin or kill
478 * its children or threads, just set TIF_MEMDIE so it can die quickly
479 */
480 if (p->flags & PF_EXITING) {
481 set_tsk_thread_flag(p, TIF_MEMDIE);
482 return 0;
483 }
484
485 task_lock(p);
486 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
487 message, task_pid_nr(p), p->comm, points);
488 task_unlock(p);
489
490 /*
491 * If any of p's children has a different mm and is eligible for kill,
492 * the one with the highest oom_badness() score is sacrificed for its
493 * parent. This attempts to lose the minimal amount of work done while
494 * still freeing memory.
495 */
496 do {
497 list_for_each_entry(child, &t->children, sibling) {
498 unsigned int child_points;
499
500 if (child->mm == p->mm)
501 continue;
502 /*
503 * oom_badness() returns 0 if the thread is unkillable
504 */
505 child_points = oom_badness(child, mem, nodemask,
506 totalpages);
507 if (child_points > victim_points) {
508 victim = child;
509 victim_points = child_points;
510 }
511 }
512 } while_each_thread(p, t);
513
514 return oom_kill_task(victim, mem);
515}
516
517/*
518 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
519 */
520static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
521 int order, const nodemask_t *nodemask)
522{
523 if (likely(!sysctl_panic_on_oom))
524 return;
525 if (sysctl_panic_on_oom != 2) {
526 /*
527 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
528 * does not panic for cpuset, mempolicy, or memcg allocation
529 * failures.
530 */
531 if (constraint != CONSTRAINT_NONE)
532 return;
533 }
534 read_lock(&tasklist_lock);
535 dump_header(NULL, gfp_mask, order, NULL, nodemask);
536 read_unlock(&tasklist_lock);
537 panic("Out of memory: %s panic_on_oom is enabled\n",
538 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
539}
540
541#ifdef CONFIG_CGROUP_MEM_RES_CTLR
542void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
543{
544 unsigned long limit;
545 unsigned int points = 0;
546 struct task_struct *p;
547
548 /*
549 * If current has a pending SIGKILL, then automatically select it. The
550 * goal is to allow it to allocate so that it may quickly exit and free
551 * its memory.
552 */
553 if (fatal_signal_pending(current)) {
554 set_thread_flag(TIF_MEMDIE);
555 return;
556 }
557
558 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
559 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
560 read_lock(&tasklist_lock);
561retry:
562 p = select_bad_process(&points, limit, mem, NULL);
563 if (!p || PTR_ERR(p) == -1UL)
564 goto out;
565
566 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
567 "Memory cgroup out of memory"))
568 goto retry;
569out:
570 read_unlock(&tasklist_lock);
571}
572#endif
573
574static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
575
576int register_oom_notifier(struct notifier_block *nb)
577{
578 return blocking_notifier_chain_register(&oom_notify_list, nb);
579}
580EXPORT_SYMBOL_GPL(register_oom_notifier);
581
582int unregister_oom_notifier(struct notifier_block *nb)
583{
584 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
585}
586EXPORT_SYMBOL_GPL(unregister_oom_notifier);
587
588/*
589 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
590 * if a parallel OOM killing is already taking place that includes a zone in
591 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
592 */
593int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
594{
595 struct zoneref *z;
596 struct zone *zone;
597 int ret = 1;
598
599 spin_lock(&zone_scan_lock);
600 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
601 if (zone_is_oom_locked(zone)) {
602 ret = 0;
603 goto out;
604 }
605 }
606
607 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
608 /*
609 * Lock each zone in the zonelist under zone_scan_lock so a
610 * parallel invocation of try_set_zonelist_oom() doesn't succeed
611 * when it shouldn't.
612 */
613 zone_set_flag(zone, ZONE_OOM_LOCKED);
614 }
615
616out:
617 spin_unlock(&zone_scan_lock);
618 return ret;
619}
620
621/*
622 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
623 * allocation attempts with zonelists containing them may now recall the OOM
624 * killer, if necessary.
625 */
626void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
627{
628 struct zoneref *z;
629 struct zone *zone;
630
631 spin_lock(&zone_scan_lock);
632 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
633 zone_clear_flag(zone, ZONE_OOM_LOCKED);
634 }
635 spin_unlock(&zone_scan_lock);
636}
637
638/*
639 * Try to acquire the oom killer lock for all system zones. Returns zero if a
640 * parallel oom killing is taking place, otherwise locks all zones and returns
641 * non-zero.
642 */
643static int try_set_system_oom(void)
644{
645 struct zone *zone;
646 int ret = 1;
647
648 spin_lock(&zone_scan_lock);
649 for_each_populated_zone(zone)
650 if (zone_is_oom_locked(zone)) {
651 ret = 0;
652 goto out;
653 }
654 for_each_populated_zone(zone)
655 zone_set_flag(zone, ZONE_OOM_LOCKED);
656out:
657 spin_unlock(&zone_scan_lock);
658 return ret;
659}
660
661/*
662 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
663 * attempts or page faults may now recall the oom killer, if necessary.
664 */
665static void clear_system_oom(void)
666{
667 struct zone *zone;
668
669 spin_lock(&zone_scan_lock);
670 for_each_populated_zone(zone)
671 zone_clear_flag(zone, ZONE_OOM_LOCKED);
672 spin_unlock(&zone_scan_lock);
673}
674
675/**
676 * out_of_memory - kill the "best" process when we run out of memory
677 * @zonelist: zonelist pointer
678 * @gfp_mask: memory allocation flags
679 * @order: amount of memory being requested as a power of 2
680 * @nodemask: nodemask passed to page allocator
681 *
682 * If we run out of memory, we have the choice between either
683 * killing a random task (bad), letting the system crash (worse)
684 * OR try to be smart about which process to kill. Note that we
685 * don't have to be perfect here, we just have to be good.
686 */
687void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
688 int order, nodemask_t *nodemask)
689{
690 const nodemask_t *mpol_mask;
691 struct task_struct *p;
692 unsigned long totalpages;
693 unsigned long freed = 0;
694 unsigned int points;
695 enum oom_constraint constraint = CONSTRAINT_NONE;
696 int killed = 0;
697
698 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
699 if (freed > 0)
700 /* Got some memory back in the last second. */
701 return;
702
703 /*
704 * If current has a pending SIGKILL, then automatically select it. The
705 * goal is to allow it to allocate so that it may quickly exit and free
706 * its memory.
707 */
708 if (fatal_signal_pending(current)) {
709 set_thread_flag(TIF_MEMDIE);
710 return;
711 }
712
713 /*
714 * Check if there were limitations on the allocation (only relevant for
715 * NUMA) that may require different handling.
716 */
717 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
718 &totalpages);
719 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
720 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
721
722 read_lock(&tasklist_lock);
723 if (sysctl_oom_kill_allocating_task &&
724 !oom_unkillable_task(current, NULL, nodemask) &&
725 current->mm && !atomic_read(¤t->mm->oom_disable_count)) {
726 /*
727 * oom_kill_process() needs tasklist_lock held. If it returns
728 * non-zero, current could not be killed so we must fallback to
729 * the tasklist scan.
730 */
731 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
732 NULL, nodemask,
733 "Out of memory (oom_kill_allocating_task)"))
734 goto out;
735 }
736
737retry:
738 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
739 if (PTR_ERR(p) == -1UL)
740 goto out;
741
742 /* Found nothing?!?! Either we hang forever, or we panic. */
743 if (!p) {
744 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
745 read_unlock(&tasklist_lock);
746 panic("Out of memory and no killable processes...\n");
747 }
748
749 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
750 nodemask, "Out of memory"))
751 goto retry;
752 killed = 1;
753out:
754 read_unlock(&tasklist_lock);
755
756 /*
757 * Give "p" a good chance of killing itself before we
758 * retry to allocate memory unless "p" is current
759 */
760 if (killed && !test_thread_flag(TIF_MEMDIE))
761 schedule_timeout_uninterruptible(1);
762}
763
764/*
765 * The pagefault handler calls here because it is out of memory, so kill a
766 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
767 * oom killing is already in progress so do nothing. If a task is found with
768 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
769 */
770void pagefault_out_of_memory(void)
771{
772 if (try_set_system_oom()) {
773 out_of_memory(NULL, 0, 0, NULL);
774 clear_system_oom();
775 }
776 if (!test_thread_flag(TIF_MEMDIE))
777 schedule_timeout_uninterruptible(1);
778}
1/*
2 * linux/mm/oom_kill.c
3 *
4 * Copyright (C) 1998,2000 Rik van Riel
5 * Thanks go out to Claus Fischer for some serious inspiration and
6 * for goading me into coding this file...
7 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
9 *
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
13 *
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
18 */
19
20#include <linux/oom.h>
21#include <linux/mm.h>
22#include <linux/err.h>
23#include <linux/gfp.h>
24#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
28#include <linux/cpuset.h>
29#include <linux/export.h>
30#include <linux/notifier.h>
31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
33#include <linux/security.h>
34#include <linux/ptrace.h>
35#include <linux/freezer.h>
36#include <linux/ftrace.h>
37#include <linux/ratelimit.h>
38#include <linux/kthread.h>
39#include <linux/init.h>
40
41#include <asm/tlb.h>
42#include "internal.h"
43
44#define CREATE_TRACE_POINTS
45#include <trace/events/oom.h>
46
47int sysctl_panic_on_oom;
48int sysctl_oom_kill_allocating_task;
49int sysctl_oom_dump_tasks = 1;
50
51DEFINE_MUTEX(oom_lock);
52
53#ifdef CONFIG_NUMA
54/**
55 * has_intersects_mems_allowed() - check task eligiblity for kill
56 * @start: task struct of which task to consider
57 * @mask: nodemask passed to page allocator for mempolicy ooms
58 *
59 * Task eligibility is determined by whether or not a candidate task, @tsk,
60 * shares the same mempolicy nodes as current if it is bound by such a policy
61 * and whether or not it has the same set of allowed cpuset nodes.
62 */
63static bool has_intersects_mems_allowed(struct task_struct *start,
64 const nodemask_t *mask)
65{
66 struct task_struct *tsk;
67 bool ret = false;
68
69 rcu_read_lock();
70 for_each_thread(start, tsk) {
71 if (mask) {
72 /*
73 * If this is a mempolicy constrained oom, tsk's
74 * cpuset is irrelevant. Only return true if its
75 * mempolicy intersects current, otherwise it may be
76 * needlessly killed.
77 */
78 ret = mempolicy_nodemask_intersects(tsk, mask);
79 } else {
80 /*
81 * This is not a mempolicy constrained oom, so only
82 * check the mems of tsk's cpuset.
83 */
84 ret = cpuset_mems_allowed_intersects(current, tsk);
85 }
86 if (ret)
87 break;
88 }
89 rcu_read_unlock();
90
91 return ret;
92}
93#else
94static bool has_intersects_mems_allowed(struct task_struct *tsk,
95 const nodemask_t *mask)
96{
97 return true;
98}
99#endif /* CONFIG_NUMA */
100
101/*
102 * The process p may have detached its own ->mm while exiting or through
103 * use_mm(), but one or more of its subthreads may still have a valid
104 * pointer. Return p, or any of its subthreads with a valid ->mm, with
105 * task_lock() held.
106 */
107struct task_struct *find_lock_task_mm(struct task_struct *p)
108{
109 struct task_struct *t;
110
111 rcu_read_lock();
112
113 for_each_thread(p, t) {
114 task_lock(t);
115 if (likely(t->mm))
116 goto found;
117 task_unlock(t);
118 }
119 t = NULL;
120found:
121 rcu_read_unlock();
122
123 return t;
124}
125
126/*
127 * order == -1 means the oom kill is required by sysrq, otherwise only
128 * for display purposes.
129 */
130static inline bool is_sysrq_oom(struct oom_control *oc)
131{
132 return oc->order == -1;
133}
134
135static inline bool is_memcg_oom(struct oom_control *oc)
136{
137 return oc->memcg != NULL;
138}
139
140/* return true if the task is not adequate as candidate victim task. */
141static bool oom_unkillable_task(struct task_struct *p,
142 struct mem_cgroup *memcg, const nodemask_t *nodemask)
143{
144 if (is_global_init(p))
145 return true;
146 if (p->flags & PF_KTHREAD)
147 return true;
148
149 /* When mem_cgroup_out_of_memory() and p is not member of the group */
150 if (memcg && !task_in_mem_cgroup(p, memcg))
151 return true;
152
153 /* p may not have freeable memory in nodemask */
154 if (!has_intersects_mems_allowed(p, nodemask))
155 return true;
156
157 return false;
158}
159
160/**
161 * oom_badness - heuristic function to determine which candidate task to kill
162 * @p: task struct of which task we should calculate
163 * @totalpages: total present RAM allowed for page allocation
164 *
165 * The heuristic for determining which task to kill is made to be as simple and
166 * predictable as possible. The goal is to return the highest value for the
167 * task consuming the most memory to avoid subsequent oom failures.
168 */
169unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
170 const nodemask_t *nodemask, unsigned long totalpages)
171{
172 long points;
173 long adj;
174
175 if (oom_unkillable_task(p, memcg, nodemask))
176 return 0;
177
178 p = find_lock_task_mm(p);
179 if (!p)
180 return 0;
181
182 /*
183 * Do not even consider tasks which are explicitly marked oom
184 * unkillable or have been already oom reaped or the are in
185 * the middle of vfork
186 */
187 adj = (long)p->signal->oom_score_adj;
188 if (adj == OOM_SCORE_ADJ_MIN ||
189 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
190 in_vfork(p)) {
191 task_unlock(p);
192 return 0;
193 }
194
195 /*
196 * The baseline for the badness score is the proportion of RAM that each
197 * task's rss, pagetable and swap space use.
198 */
199 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
200 atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
201 task_unlock(p);
202
203 /*
204 * Root processes get 3% bonus, just like the __vm_enough_memory()
205 * implementation used by LSMs.
206 */
207 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
208 points -= (points * 3) / 100;
209
210 /* Normalize to oom_score_adj units */
211 adj *= totalpages / 1000;
212 points += adj;
213
214 /*
215 * Never return 0 for an eligible task regardless of the root bonus and
216 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
217 */
218 return points > 0 ? points : 1;
219}
220
221enum oom_constraint {
222 CONSTRAINT_NONE,
223 CONSTRAINT_CPUSET,
224 CONSTRAINT_MEMORY_POLICY,
225 CONSTRAINT_MEMCG,
226};
227
228/*
229 * Determine the type of allocation constraint.
230 */
231static enum oom_constraint constrained_alloc(struct oom_control *oc)
232{
233 struct zone *zone;
234 struct zoneref *z;
235 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
236 bool cpuset_limited = false;
237 int nid;
238
239 if (is_memcg_oom(oc)) {
240 oc->totalpages = mem_cgroup_get_limit(oc->memcg) ?: 1;
241 return CONSTRAINT_MEMCG;
242 }
243
244 /* Default to all available memory */
245 oc->totalpages = totalram_pages + total_swap_pages;
246
247 if (!IS_ENABLED(CONFIG_NUMA))
248 return CONSTRAINT_NONE;
249
250 if (!oc->zonelist)
251 return CONSTRAINT_NONE;
252 /*
253 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
254 * to kill current.We have to random task kill in this case.
255 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
256 */
257 if (oc->gfp_mask & __GFP_THISNODE)
258 return CONSTRAINT_NONE;
259
260 /*
261 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
262 * the page allocator means a mempolicy is in effect. Cpuset policy
263 * is enforced in get_page_from_freelist().
264 */
265 if (oc->nodemask &&
266 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
267 oc->totalpages = total_swap_pages;
268 for_each_node_mask(nid, *oc->nodemask)
269 oc->totalpages += node_spanned_pages(nid);
270 return CONSTRAINT_MEMORY_POLICY;
271 }
272
273 /* Check this allocation failure is caused by cpuset's wall function */
274 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
275 high_zoneidx, oc->nodemask)
276 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
277 cpuset_limited = true;
278
279 if (cpuset_limited) {
280 oc->totalpages = total_swap_pages;
281 for_each_node_mask(nid, cpuset_current_mems_allowed)
282 oc->totalpages += node_spanned_pages(nid);
283 return CONSTRAINT_CPUSET;
284 }
285 return CONSTRAINT_NONE;
286}
287
288static int oom_evaluate_task(struct task_struct *task, void *arg)
289{
290 struct oom_control *oc = arg;
291 unsigned long points;
292
293 if (oom_unkillable_task(task, NULL, oc->nodemask))
294 goto next;
295
296 /*
297 * This task already has access to memory reserves and is being killed.
298 * Don't allow any other task to have access to the reserves unless
299 * the task has MMF_OOM_SKIP because chances that it would release
300 * any memory is quite low.
301 */
302 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
303 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
304 goto next;
305 goto abort;
306 }
307
308 /*
309 * If task is allocating a lot of memory and has been marked to be
310 * killed first if it triggers an oom, then select it.
311 */
312 if (oom_task_origin(task)) {
313 points = ULONG_MAX;
314 goto select;
315 }
316
317 points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
318 if (!points || points < oc->chosen_points)
319 goto next;
320
321 /* Prefer thread group leaders for display purposes */
322 if (points == oc->chosen_points && thread_group_leader(oc->chosen))
323 goto next;
324select:
325 if (oc->chosen)
326 put_task_struct(oc->chosen);
327 get_task_struct(task);
328 oc->chosen = task;
329 oc->chosen_points = points;
330next:
331 return 0;
332abort:
333 if (oc->chosen)
334 put_task_struct(oc->chosen);
335 oc->chosen = (void *)-1UL;
336 return 1;
337}
338
339/*
340 * Simple selection loop. We choose the process with the highest number of
341 * 'points'. In case scan was aborted, oc->chosen is set to -1.
342 */
343static void select_bad_process(struct oom_control *oc)
344{
345 if (is_memcg_oom(oc))
346 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
347 else {
348 struct task_struct *p;
349
350 rcu_read_lock();
351 for_each_process(p)
352 if (oom_evaluate_task(p, oc))
353 break;
354 rcu_read_unlock();
355 }
356
357 oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages;
358}
359
360/**
361 * dump_tasks - dump current memory state of all system tasks
362 * @memcg: current's memory controller, if constrained
363 * @nodemask: nodemask passed to page allocator for mempolicy ooms
364 *
365 * Dumps the current memory state of all eligible tasks. Tasks not in the same
366 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
367 * are not shown.
368 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
369 * swapents, oom_score_adj value, and name.
370 */
371static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
372{
373 struct task_struct *p;
374 struct task_struct *task;
375
376 pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n");
377 rcu_read_lock();
378 for_each_process(p) {
379 if (oom_unkillable_task(p, memcg, nodemask))
380 continue;
381
382 task = find_lock_task_mm(p);
383 if (!task) {
384 /*
385 * This is a kthread or all of p's threads have already
386 * detached their mm's. There's no need to report
387 * them; they can't be oom killed anyway.
388 */
389 continue;
390 }
391
392 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n",
393 task->pid, from_kuid(&init_user_ns, task_uid(task)),
394 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
395 atomic_long_read(&task->mm->nr_ptes),
396 mm_nr_pmds(task->mm),
397 get_mm_counter(task->mm, MM_SWAPENTS),
398 task->signal->oom_score_adj, task->comm);
399 task_unlock(task);
400 }
401 rcu_read_unlock();
402}
403
404static void dump_header(struct oom_control *oc, struct task_struct *p)
405{
406 nodemask_t *nm = (oc->nodemask) ? oc->nodemask : &cpuset_current_mems_allowed;
407
408 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n",
409 current->comm, oc->gfp_mask, &oc->gfp_mask,
410 nodemask_pr_args(nm), oc->order,
411 current->signal->oom_score_adj);
412 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
413 pr_warn("COMPACTION is disabled!!!\n");
414
415 cpuset_print_current_mems_allowed();
416 dump_stack();
417 if (oc->memcg)
418 mem_cgroup_print_oom_info(oc->memcg, p);
419 else
420 show_mem(SHOW_MEM_FILTER_NODES);
421 if (sysctl_oom_dump_tasks)
422 dump_tasks(oc->memcg, oc->nodemask);
423}
424
425/*
426 * Number of OOM victims in flight
427 */
428static atomic_t oom_victims = ATOMIC_INIT(0);
429static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
430
431static bool oom_killer_disabled __read_mostly;
432
433#define K(x) ((x) << (PAGE_SHIFT-10))
434
435/*
436 * task->mm can be NULL if the task is the exited group leader. So to
437 * determine whether the task is using a particular mm, we examine all the
438 * task's threads: if one of those is using this mm then this task was also
439 * using it.
440 */
441bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
442{
443 struct task_struct *t;
444
445 for_each_thread(p, t) {
446 struct mm_struct *t_mm = READ_ONCE(t->mm);
447 if (t_mm)
448 return t_mm == mm;
449 }
450 return false;
451}
452
453
454#ifdef CONFIG_MMU
455/*
456 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
457 * victim (if that is possible) to help the OOM killer to move on.
458 */
459static struct task_struct *oom_reaper_th;
460static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
461static struct task_struct *oom_reaper_list;
462static DEFINE_SPINLOCK(oom_reaper_lock);
463
464static bool __oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
465{
466 struct mmu_gather tlb;
467 struct vm_area_struct *vma;
468 struct zap_details details = {.check_swap_entries = true,
469 .ignore_dirty = true};
470 bool ret = true;
471
472 /*
473 * We have to make sure to not race with the victim exit path
474 * and cause premature new oom victim selection:
475 * __oom_reap_task_mm exit_mm
476 * mmget_not_zero
477 * mmput
478 * atomic_dec_and_test
479 * exit_oom_victim
480 * [...]
481 * out_of_memory
482 * select_bad_process
483 * # no TIF_MEMDIE task selects new victim
484 * unmap_page_range # frees some memory
485 */
486 mutex_lock(&oom_lock);
487
488 if (!down_read_trylock(&mm->mmap_sem)) {
489 ret = false;
490 goto unlock_oom;
491 }
492
493 /*
494 * increase mm_users only after we know we will reap something so
495 * that the mmput_async is called only when we have reaped something
496 * and delayed __mmput doesn't matter that much
497 */
498 if (!mmget_not_zero(mm)) {
499 up_read(&mm->mmap_sem);
500 goto unlock_oom;
501 }
502
503 /*
504 * Tell all users of get_user/copy_from_user etc... that the content
505 * is no longer stable. No barriers really needed because unmapping
506 * should imply barriers already and the reader would hit a page fault
507 * if it stumbled over a reaped memory.
508 */
509 set_bit(MMF_UNSTABLE, &mm->flags);
510
511 tlb_gather_mmu(&tlb, mm, 0, -1);
512 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
513 if (is_vm_hugetlb_page(vma))
514 continue;
515
516 /*
517 * mlocked VMAs require explicit munlocking before unmap.
518 * Let's keep it simple here and skip such VMAs.
519 */
520 if (vma->vm_flags & VM_LOCKED)
521 continue;
522
523 /*
524 * Only anonymous pages have a good chance to be dropped
525 * without additional steps which we cannot afford as we
526 * are OOM already.
527 *
528 * We do not even care about fs backed pages because all
529 * which are reclaimable have already been reclaimed and
530 * we do not want to block exit_mmap by keeping mm ref
531 * count elevated without a good reason.
532 */
533 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED))
534 unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end,
535 &details);
536 }
537 tlb_finish_mmu(&tlb, 0, -1);
538 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
539 task_pid_nr(tsk), tsk->comm,
540 K(get_mm_counter(mm, MM_ANONPAGES)),
541 K(get_mm_counter(mm, MM_FILEPAGES)),
542 K(get_mm_counter(mm, MM_SHMEMPAGES)));
543 up_read(&mm->mmap_sem);
544
545 /*
546 * Drop our reference but make sure the mmput slow path is called from a
547 * different context because we shouldn't risk we get stuck there and
548 * put the oom_reaper out of the way.
549 */
550 mmput_async(mm);
551unlock_oom:
552 mutex_unlock(&oom_lock);
553 return ret;
554}
555
556#define MAX_OOM_REAP_RETRIES 10
557static void oom_reap_task(struct task_struct *tsk)
558{
559 int attempts = 0;
560 struct mm_struct *mm = tsk->signal->oom_mm;
561
562 /* Retry the down_read_trylock(mmap_sem) a few times */
563 while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task_mm(tsk, mm))
564 schedule_timeout_idle(HZ/10);
565
566 if (attempts <= MAX_OOM_REAP_RETRIES)
567 goto done;
568
569
570 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
571 task_pid_nr(tsk), tsk->comm);
572 debug_show_all_locks();
573
574done:
575 tsk->oom_reaper_list = NULL;
576
577 /*
578 * Hide this mm from OOM killer because it has been either reaped or
579 * somebody can't call up_write(mmap_sem).
580 */
581 set_bit(MMF_OOM_SKIP, &mm->flags);
582
583 /* Drop a reference taken by wake_oom_reaper */
584 put_task_struct(tsk);
585}
586
587static int oom_reaper(void *unused)
588{
589 while (true) {
590 struct task_struct *tsk = NULL;
591
592 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
593 spin_lock(&oom_reaper_lock);
594 if (oom_reaper_list != NULL) {
595 tsk = oom_reaper_list;
596 oom_reaper_list = tsk->oom_reaper_list;
597 }
598 spin_unlock(&oom_reaper_lock);
599
600 if (tsk)
601 oom_reap_task(tsk);
602 }
603
604 return 0;
605}
606
607static void wake_oom_reaper(struct task_struct *tsk)
608{
609 if (!oom_reaper_th)
610 return;
611
612 /* tsk is already queued? */
613 if (tsk == oom_reaper_list || tsk->oom_reaper_list)
614 return;
615
616 get_task_struct(tsk);
617
618 spin_lock(&oom_reaper_lock);
619 tsk->oom_reaper_list = oom_reaper_list;
620 oom_reaper_list = tsk;
621 spin_unlock(&oom_reaper_lock);
622 wake_up(&oom_reaper_wait);
623}
624
625static int __init oom_init(void)
626{
627 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
628 if (IS_ERR(oom_reaper_th)) {
629 pr_err("Unable to start OOM reaper %ld. Continuing regardless\n",
630 PTR_ERR(oom_reaper_th));
631 oom_reaper_th = NULL;
632 }
633 return 0;
634}
635subsys_initcall(oom_init)
636#else
637static inline void wake_oom_reaper(struct task_struct *tsk)
638{
639}
640#endif /* CONFIG_MMU */
641
642/**
643 * mark_oom_victim - mark the given task as OOM victim
644 * @tsk: task to mark
645 *
646 * Has to be called with oom_lock held and never after
647 * oom has been disabled already.
648 *
649 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
650 * under task_lock or operate on the current).
651 */
652static void mark_oom_victim(struct task_struct *tsk)
653{
654 struct mm_struct *mm = tsk->mm;
655
656 WARN_ON(oom_killer_disabled);
657 /* OOM killer might race with memcg OOM */
658 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
659 return;
660
661 /* oom_mm is bound to the signal struct life time. */
662 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
663 atomic_inc(&tsk->signal->oom_mm->mm_count);
664
665 /*
666 * Make sure that the task is woken up from uninterruptible sleep
667 * if it is frozen because OOM killer wouldn't be able to free
668 * any memory and livelock. freezing_slow_path will tell the freezer
669 * that TIF_MEMDIE tasks should be ignored.
670 */
671 __thaw_task(tsk);
672 atomic_inc(&oom_victims);
673}
674
675/**
676 * exit_oom_victim - note the exit of an OOM victim
677 */
678void exit_oom_victim(void)
679{
680 clear_thread_flag(TIF_MEMDIE);
681
682 if (!atomic_dec_return(&oom_victims))
683 wake_up_all(&oom_victims_wait);
684}
685
686/**
687 * oom_killer_enable - enable OOM killer
688 */
689void oom_killer_enable(void)
690{
691 oom_killer_disabled = false;
692}
693
694/**
695 * oom_killer_disable - disable OOM killer
696 * @timeout: maximum timeout to wait for oom victims in jiffies
697 *
698 * Forces all page allocations to fail rather than trigger OOM killer.
699 * Will block and wait until all OOM victims are killed or the given
700 * timeout expires.
701 *
702 * The function cannot be called when there are runnable user tasks because
703 * the userspace would see unexpected allocation failures as a result. Any
704 * new usage of this function should be consulted with MM people.
705 *
706 * Returns true if successful and false if the OOM killer cannot be
707 * disabled.
708 */
709bool oom_killer_disable(signed long timeout)
710{
711 signed long ret;
712
713 /*
714 * Make sure to not race with an ongoing OOM killer. Check that the
715 * current is not killed (possibly due to sharing the victim's memory).
716 */
717 if (mutex_lock_killable(&oom_lock))
718 return false;
719 oom_killer_disabled = true;
720 mutex_unlock(&oom_lock);
721
722 ret = wait_event_interruptible_timeout(oom_victims_wait,
723 !atomic_read(&oom_victims), timeout);
724 if (ret <= 0) {
725 oom_killer_enable();
726 return false;
727 }
728
729 return true;
730}
731
732static inline bool __task_will_free_mem(struct task_struct *task)
733{
734 struct signal_struct *sig = task->signal;
735
736 /*
737 * A coredumping process may sleep for an extended period in exit_mm(),
738 * so the oom killer cannot assume that the process will promptly exit
739 * and release memory.
740 */
741 if (sig->flags & SIGNAL_GROUP_COREDUMP)
742 return false;
743
744 if (sig->flags & SIGNAL_GROUP_EXIT)
745 return true;
746
747 if (thread_group_empty(task) && (task->flags & PF_EXITING))
748 return true;
749
750 return false;
751}
752
753/*
754 * Checks whether the given task is dying or exiting and likely to
755 * release its address space. This means that all threads and processes
756 * sharing the same mm have to be killed or exiting.
757 * Caller has to make sure that task->mm is stable (hold task_lock or
758 * it operates on the current).
759 */
760static bool task_will_free_mem(struct task_struct *task)
761{
762 struct mm_struct *mm = task->mm;
763 struct task_struct *p;
764 bool ret = true;
765
766 /*
767 * Skip tasks without mm because it might have passed its exit_mm and
768 * exit_oom_victim. oom_reaper could have rescued that but do not rely
769 * on that for now. We can consider find_lock_task_mm in future.
770 */
771 if (!mm)
772 return false;
773
774 if (!__task_will_free_mem(task))
775 return false;
776
777 /*
778 * This task has already been drained by the oom reaper so there are
779 * only small chances it will free some more
780 */
781 if (test_bit(MMF_OOM_SKIP, &mm->flags))
782 return false;
783
784 if (atomic_read(&mm->mm_users) <= 1)
785 return true;
786
787 /*
788 * Make sure that all tasks which share the mm with the given tasks
789 * are dying as well to make sure that a) nobody pins its mm and
790 * b) the task is also reapable by the oom reaper.
791 */
792 rcu_read_lock();
793 for_each_process(p) {
794 if (!process_shares_mm(p, mm))
795 continue;
796 if (same_thread_group(task, p))
797 continue;
798 ret = __task_will_free_mem(p);
799 if (!ret)
800 break;
801 }
802 rcu_read_unlock();
803
804 return ret;
805}
806
807static void oom_kill_process(struct oom_control *oc, const char *message)
808{
809 struct task_struct *p = oc->chosen;
810 unsigned int points = oc->chosen_points;
811 struct task_struct *victim = p;
812 struct task_struct *child;
813 struct task_struct *t;
814 struct mm_struct *mm;
815 unsigned int victim_points = 0;
816 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
817 DEFAULT_RATELIMIT_BURST);
818 bool can_oom_reap = true;
819
820 /*
821 * If the task is already exiting, don't alarm the sysadmin or kill
822 * its children or threads, just set TIF_MEMDIE so it can die quickly
823 */
824 task_lock(p);
825 if (task_will_free_mem(p)) {
826 mark_oom_victim(p);
827 wake_oom_reaper(p);
828 task_unlock(p);
829 put_task_struct(p);
830 return;
831 }
832 task_unlock(p);
833
834 if (__ratelimit(&oom_rs))
835 dump_header(oc, p);
836
837 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
838 message, task_pid_nr(p), p->comm, points);
839
840 /*
841 * If any of p's children has a different mm and is eligible for kill,
842 * the one with the highest oom_badness() score is sacrificed for its
843 * parent. This attempts to lose the minimal amount of work done while
844 * still freeing memory.
845 */
846 read_lock(&tasklist_lock);
847 for_each_thread(p, t) {
848 list_for_each_entry(child, &t->children, sibling) {
849 unsigned int child_points;
850
851 if (process_shares_mm(child, p->mm))
852 continue;
853 /*
854 * oom_badness() returns 0 if the thread is unkillable
855 */
856 child_points = oom_badness(child,
857 oc->memcg, oc->nodemask, oc->totalpages);
858 if (child_points > victim_points) {
859 put_task_struct(victim);
860 victim = child;
861 victim_points = child_points;
862 get_task_struct(victim);
863 }
864 }
865 }
866 read_unlock(&tasklist_lock);
867
868 p = find_lock_task_mm(victim);
869 if (!p) {
870 put_task_struct(victim);
871 return;
872 } else if (victim != p) {
873 get_task_struct(p);
874 put_task_struct(victim);
875 victim = p;
876 }
877
878 /* Get a reference to safely compare mm after task_unlock(victim) */
879 mm = victim->mm;
880 atomic_inc(&mm->mm_count);
881 /*
882 * We should send SIGKILL before setting TIF_MEMDIE in order to prevent
883 * the OOM victim from depleting the memory reserves from the user
884 * space under its control.
885 */
886 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
887 mark_oom_victim(victim);
888 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
889 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
890 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
891 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
892 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
893 task_unlock(victim);
894
895 /*
896 * Kill all user processes sharing victim->mm in other thread groups, if
897 * any. They don't get access to memory reserves, though, to avoid
898 * depletion of all memory. This prevents mm->mmap_sem livelock when an
899 * oom killed thread cannot exit because it requires the semaphore and
900 * its contended by another thread trying to allocate memory itself.
901 * That thread will now get access to memory reserves since it has a
902 * pending fatal signal.
903 */
904 rcu_read_lock();
905 for_each_process(p) {
906 if (!process_shares_mm(p, mm))
907 continue;
908 if (same_thread_group(p, victim))
909 continue;
910 if (is_global_init(p)) {
911 can_oom_reap = false;
912 set_bit(MMF_OOM_SKIP, &mm->flags);
913 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
914 task_pid_nr(victim), victim->comm,
915 task_pid_nr(p), p->comm);
916 continue;
917 }
918 /*
919 * No use_mm() user needs to read from the userspace so we are
920 * ok to reap it.
921 */
922 if (unlikely(p->flags & PF_KTHREAD))
923 continue;
924 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
925 }
926 rcu_read_unlock();
927
928 if (can_oom_reap)
929 wake_oom_reaper(victim);
930
931 mmdrop(mm);
932 put_task_struct(victim);
933}
934#undef K
935
936/*
937 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
938 */
939static void check_panic_on_oom(struct oom_control *oc,
940 enum oom_constraint constraint)
941{
942 if (likely(!sysctl_panic_on_oom))
943 return;
944 if (sysctl_panic_on_oom != 2) {
945 /*
946 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
947 * does not panic for cpuset, mempolicy, or memcg allocation
948 * failures.
949 */
950 if (constraint != CONSTRAINT_NONE)
951 return;
952 }
953 /* Do not panic for oom kills triggered by sysrq */
954 if (is_sysrq_oom(oc))
955 return;
956 dump_header(oc, NULL);
957 panic("Out of memory: %s panic_on_oom is enabled\n",
958 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
959}
960
961static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
962
963int register_oom_notifier(struct notifier_block *nb)
964{
965 return blocking_notifier_chain_register(&oom_notify_list, nb);
966}
967EXPORT_SYMBOL_GPL(register_oom_notifier);
968
969int unregister_oom_notifier(struct notifier_block *nb)
970{
971 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
972}
973EXPORT_SYMBOL_GPL(unregister_oom_notifier);
974
975/**
976 * out_of_memory - kill the "best" process when we run out of memory
977 * @oc: pointer to struct oom_control
978 *
979 * If we run out of memory, we have the choice between either
980 * killing a random task (bad), letting the system crash (worse)
981 * OR try to be smart about which process to kill. Note that we
982 * don't have to be perfect here, we just have to be good.
983 */
984bool out_of_memory(struct oom_control *oc)
985{
986 unsigned long freed = 0;
987 enum oom_constraint constraint = CONSTRAINT_NONE;
988
989 if (oom_killer_disabled)
990 return false;
991
992 if (!is_memcg_oom(oc)) {
993 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
994 if (freed > 0)
995 /* Got some memory back in the last second. */
996 return true;
997 }
998
999 /*
1000 * If current has a pending SIGKILL or is exiting, then automatically
1001 * select it. The goal is to allow it to allocate so that it may
1002 * quickly exit and free its memory.
1003 */
1004 if (task_will_free_mem(current)) {
1005 mark_oom_victim(current);
1006 wake_oom_reaper(current);
1007 return true;
1008 }
1009
1010 /*
1011 * The OOM killer does not compensate for IO-less reclaim.
1012 * pagefault_out_of_memory lost its gfp context so we have to
1013 * make sure exclude 0 mask - all other users should have at least
1014 * ___GFP_DIRECT_RECLAIM to get here.
1015 */
1016 if (oc->gfp_mask && !(oc->gfp_mask & (__GFP_FS|__GFP_NOFAIL)))
1017 return true;
1018
1019 /*
1020 * Check if there were limitations on the allocation (only relevant for
1021 * NUMA and memcg) that may require different handling.
1022 */
1023 constraint = constrained_alloc(oc);
1024 if (constraint != CONSTRAINT_MEMORY_POLICY)
1025 oc->nodemask = NULL;
1026 check_panic_on_oom(oc, constraint);
1027
1028 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1029 current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) &&
1030 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1031 get_task_struct(current);
1032 oc->chosen = current;
1033 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1034 return true;
1035 }
1036
1037 select_bad_process(oc);
1038 /* Found nothing?!?! Either we hang forever, or we panic. */
1039 if (!oc->chosen && !is_sysrq_oom(oc) && !is_memcg_oom(oc)) {
1040 dump_header(oc, NULL);
1041 panic("Out of memory and no killable processes...\n");
1042 }
1043 if (oc->chosen && oc->chosen != (void *)-1UL) {
1044 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1045 "Memory cgroup out of memory");
1046 /*
1047 * Give the killed process a good chance to exit before trying
1048 * to allocate memory again.
1049 */
1050 schedule_timeout_killable(1);
1051 }
1052 return !!oc->chosen;
1053}
1054
1055/*
1056 * The pagefault handler calls here because it is out of memory, so kill a
1057 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1058 * killing is already in progress so do nothing.
1059 */
1060void pagefault_out_of_memory(void)
1061{
1062 struct oom_control oc = {
1063 .zonelist = NULL,
1064 .nodemask = NULL,
1065 .memcg = NULL,
1066 .gfp_mask = 0,
1067 .order = 0,
1068 };
1069
1070 if (mem_cgroup_oom_synchronize(true))
1071 return;
1072
1073 if (!mutex_trylock(&oom_lock))
1074 return;
1075 out_of_memory(&oc);
1076 mutex_unlock(&oom_lock);
1077}