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 *
8 * The routines in this file are used to kill a process when
9 * we're seriously out of memory. This gets called from __alloc_pages()
10 * in mm/page_alloc.c when we really run out of memory.
11 *
12 * Since we won't call these routines often (on a well-configured
13 * machine) this file will double as a 'coding guide' and a signpost
14 * for newbie kernel hackers. It features several pointers to major
15 * kernel subsystems and hints as to where to find out what things do.
16 */
17
18 #include <linux/oom.h>
19 #include <linux/mm.h>
20 #include <linux/err.h>
21 #include <linux/sched.h>
22 #include <linux/swap.h>
23 #include <linux/timex.h>
24 #include <linux/jiffies.h>
25 #include <linux/cpuset.h>
26 #include <linux/module.h>
27 #include <linux/notifier.h>
28 #include <linux/memcontrol.h>
29 #include <linux/security.h>
30
31 int sysctl_panic_on_oom;
32 int sysctl_oom_kill_allocating_task;
33 int sysctl_oom_dump_tasks;
34 static DEFINE_SPINLOCK(zone_scan_lock);
35 /* #define DEBUG */
36
37 /**
38 * badness - calculate a numeric value for how bad this task has been
39 * @p: task struct of which task we should calculate
40 * @uptime: current uptime in seconds
41 *
42 * The formula used is relatively simple and documented inline in the
43 * function. The main rationale is that we want to select a good task
44 * to kill when we run out of memory.
45 *
46 * Good in this context means that:
47 * 1) we lose the minimum amount of work done
48 * 2) we recover a large amount of memory
49 * 3) we don't kill anything innocent of eating tons of memory
50 * 4) we want to kill the minimum amount of processes (one)
51 * 5) we try to kill the process the user expects us to kill, this
52 * algorithm has been meticulously tuned to meet the principle
53 * of least surprise ... (be careful when you change it)
54 */
55
badness(struct task_struct * p,unsigned long uptime)56 unsigned long badness(struct task_struct *p, unsigned long uptime)
57 {
58 unsigned long points, cpu_time, run_time, s;
59 struct mm_struct *mm;
60 struct task_struct *child;
61
62 task_lock(p);
63 mm = p->mm;
64 if (!mm) {
65 task_unlock(p);
66 return 0;
67 }
68
69 /*
70 * The memory size of the process is the basis for the badness.
71 */
72 points = mm->total_vm;
73
74 /*
75 * After this unlock we can no longer dereference local variable `mm'
76 */
77 task_unlock(p);
78
79 /*
80 * swapoff can easily use up all memory, so kill those first.
81 */
82 if (p->flags & PF_SWAPOFF)
83 return ULONG_MAX;
84
85 /*
86 * Processes which fork a lot of child processes are likely
87 * a good choice. We add half the vmsize of the children if they
88 * have an own mm. This prevents forking servers to flood the
89 * machine with an endless amount of children. In case a single
90 * child is eating the vast majority of memory, adding only half
91 * to the parents will make the child our kill candidate of choice.
92 */
93 list_for_each_entry(child, &p->children, sibling) {
94 task_lock(child);
95 if (child->mm != mm && child->mm)
96 points += child->mm->total_vm/2 + 1;
97 task_unlock(child);
98 }
99
100 /*
101 * CPU time is in tens of seconds and run time is in thousands
102 * of seconds. There is no particular reason for this other than
103 * that it turned out to work very well in practice.
104 */
105 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
106 >> (SHIFT_HZ + 3);
107
108 if (uptime >= p->start_time.tv_sec)
109 run_time = (uptime - p->start_time.tv_sec) >> 10;
110 else
111 run_time = 0;
112
113 s = int_sqrt(cpu_time);
114 if (s)
115 points /= s;
116 s = int_sqrt(int_sqrt(run_time));
117 if (s)
118 points /= s;
119
120 /*
121 * Niced processes are most likely less important, so double
122 * their badness points.
123 */
124 if (task_nice(p) > 0)
125 points *= 2;
126
127 /*
128 * Superuser processes are usually more important, so we make it
129 * less likely that we kill those.
130 */
131 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
132 has_capability_noaudit(p, CAP_SYS_RESOURCE))
133 points /= 4;
134
135 /*
136 * We don't want to kill a process with direct hardware access.
137 * Not only could that mess up the hardware, but usually users
138 * tend to only have this flag set on applications they think
139 * of as important.
140 */
141 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
142 points /= 4;
143
144 /*
145 * If p's nodes don't overlap ours, it may still help to kill p
146 * because p may have allocated or otherwise mapped memory on
147 * this node before. However it will be less likely.
148 */
149 if (!cpuset_mems_allowed_intersects(current, p))
150 points /= 8;
151
152 /*
153 * Adjust the score by oomkilladj.
154 */
155 if (p->oomkilladj) {
156 if (p->oomkilladj > 0) {
157 if (!points)
158 points = 1;
159 points <<= p->oomkilladj;
160 } else
161 points >>= -(p->oomkilladj);
162 }
163
164 #ifdef DEBUG
165 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
166 p->pid, p->comm, points);
167 #endif
168 return points;
169 }
170
171 /*
172 * Determine the type of allocation constraint.
173 */
constrained_alloc(struct zonelist * zonelist,gfp_t gfp_mask)174 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
175 gfp_t gfp_mask)
176 {
177 #ifdef CONFIG_NUMA
178 struct zone *zone;
179 struct zoneref *z;
180 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
181 nodemask_t nodes = node_states[N_HIGH_MEMORY];
182
183 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
184 if (cpuset_zone_allowed_softwall(zone, gfp_mask))
185 node_clear(zone_to_nid(zone), nodes);
186 else
187 return CONSTRAINT_CPUSET;
188
189 if (!nodes_empty(nodes))
190 return CONSTRAINT_MEMORY_POLICY;
191 #endif
192
193 return CONSTRAINT_NONE;
194 }
195
196 /*
197 * Simple selection loop. We chose the process with the highest
198 * number of 'points'. We expect the caller will lock the tasklist.
199 *
200 * (not docbooked, we don't want this one cluttering up the manual)
201 */
select_bad_process(unsigned long * ppoints,struct mem_cgroup * mem)202 static struct task_struct *select_bad_process(unsigned long *ppoints,
203 struct mem_cgroup *mem)
204 {
205 struct task_struct *g, *p;
206 struct task_struct *chosen = NULL;
207 struct timespec uptime;
208 *ppoints = 0;
209
210 do_posix_clock_monotonic_gettime(&uptime);
211 do_each_thread(g, p) {
212 unsigned long points;
213
214 /*
215 * skip kernel threads and tasks which have already released
216 * their mm.
217 */
218 if (!p->mm)
219 continue;
220 /* skip the init task */
221 if (is_global_init(p))
222 continue;
223 if (mem && !task_in_mem_cgroup(p, mem))
224 continue;
225
226 /*
227 * This task already has access to memory reserves and is
228 * being killed. Don't allow any other task access to the
229 * memory reserve.
230 *
231 * Note: this may have a chance of deadlock if it gets
232 * blocked waiting for another task which itself is waiting
233 * for memory. Is there a better alternative?
234 */
235 if (test_tsk_thread_flag(p, TIF_MEMDIE))
236 return ERR_PTR(-1UL);
237
238 /*
239 * This is in the process of releasing memory so wait for it
240 * to finish before killing some other task by mistake.
241 *
242 * However, if p is the current task, we allow the 'kill' to
243 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
244 * which will allow it to gain access to memory reserves in
245 * the process of exiting and releasing its resources.
246 * Otherwise we could get an easy OOM deadlock.
247 */
248 if (p->flags & PF_EXITING) {
249 if (p != current)
250 return ERR_PTR(-1UL);
251
252 chosen = p;
253 *ppoints = ULONG_MAX;
254 }
255
256 if (p->oomkilladj == OOM_DISABLE)
257 continue;
258
259 points = badness(p, uptime.tv_sec);
260 if (points > *ppoints || !chosen) {
261 chosen = p;
262 *ppoints = points;
263 }
264 } while_each_thread(g, p);
265
266 return chosen;
267 }
268
269 /**
270 * dump_tasks - dump current memory state of all system tasks
271 * @mem: target memory controller
272 *
273 * Dumps the current memory state of all system tasks, excluding kernel threads.
274 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
275 * score, and name.
276 *
277 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
278 * shown.
279 *
280 * Call with tasklist_lock read-locked.
281 */
dump_tasks(const struct mem_cgroup * mem)282 static void dump_tasks(const struct mem_cgroup *mem)
283 {
284 struct task_struct *g, *p;
285
286 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
287 "name\n");
288 do_each_thread(g, p) {
289 /*
290 * total_vm and rss sizes do not exist for tasks with a
291 * detached mm so there's no need to report them.
292 */
293 if (!p->mm)
294 continue;
295 if (mem && !task_in_mem_cgroup(p, mem))
296 continue;
297 if (!thread_group_leader(p))
298 continue;
299
300 task_lock(p);
301 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
302 p->pid, __task_cred(p)->uid, p->tgid,
303 p->mm->total_vm, get_mm_rss(p->mm), (int)task_cpu(p),
304 p->oomkilladj, p->comm);
305 task_unlock(p);
306 } while_each_thread(g, p);
307 }
308
309 /*
310 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
311 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
312 * set.
313 */
__oom_kill_task(struct task_struct * p,int verbose)314 static void __oom_kill_task(struct task_struct *p, int verbose)
315 {
316 if (is_global_init(p)) {
317 WARN_ON(1);
318 printk(KERN_WARNING "tried to kill init!\n");
319 return;
320 }
321
322 if (!p->mm) {
323 WARN_ON(1);
324 printk(KERN_WARNING "tried to kill an mm-less task!\n");
325 return;
326 }
327
328 if (verbose)
329 printk(KERN_ERR "Killed process %d (%s)\n",
330 task_pid_nr(p), p->comm);
331
332 /*
333 * We give our sacrificial lamb high priority and access to
334 * all the memory it needs. That way it should be able to
335 * exit() and clear out its resources quickly...
336 */
337 p->rt.time_slice = HZ;
338 set_tsk_thread_flag(p, TIF_MEMDIE);
339
340 force_sig(SIGKILL, p);
341 }
342
oom_kill_task(struct task_struct * p)343 static int oom_kill_task(struct task_struct *p)
344 {
345 struct mm_struct *mm;
346 struct task_struct *g, *q;
347
348 mm = p->mm;
349
350 /* WARNING: mm may not be dereferenced since we did not obtain its
351 * value from get_task_mm(p). This is OK since all we need to do is
352 * compare mm to q->mm below.
353 *
354 * Furthermore, even if mm contains a non-NULL value, p->mm may
355 * change to NULL at any time since we do not hold task_lock(p).
356 * However, this is of no concern to us.
357 */
358
359 if (mm == NULL)
360 return 1;
361
362 /*
363 * Don't kill the process if any threads are set to OOM_DISABLE
364 */
365 do_each_thread(g, q) {
366 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
367 return 1;
368 } while_each_thread(g, q);
369
370 __oom_kill_task(p, 1);
371
372 /*
373 * kill all processes that share the ->mm (i.e. all threads),
374 * but are in a different thread group. Don't let them have access
375 * to memory reserves though, otherwise we might deplete all memory.
376 */
377 do_each_thread(g, q) {
378 if (q->mm == mm && !same_thread_group(q, p))
379 force_sig(SIGKILL, q);
380 } while_each_thread(g, q);
381
382 return 0;
383 }
384
oom_kill_process(struct task_struct * p,gfp_t gfp_mask,int order,unsigned long points,struct mem_cgroup * mem,const char * message)385 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
386 unsigned long points, struct mem_cgroup *mem,
387 const char *message)
388 {
389 struct task_struct *c;
390
391 if (printk_ratelimit()) {
392 printk(KERN_WARNING "%s invoked oom-killer: "
393 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
394 current->comm, gfp_mask, order, current->oomkilladj);
395 task_lock(current);
396 cpuset_print_task_mems_allowed(current);
397 task_unlock(current);
398 dump_stack();
399 show_mem();
400 if (sysctl_oom_dump_tasks)
401 dump_tasks(mem);
402 }
403
404 /*
405 * If the task is already exiting, don't alarm the sysadmin or kill
406 * its children or threads, just set TIF_MEMDIE so it can die quickly
407 */
408 if (p->flags & PF_EXITING) {
409 __oom_kill_task(p, 0);
410 return 0;
411 }
412
413 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
414 message, task_pid_nr(p), p->comm, points);
415
416 /* Try to kill a child first */
417 list_for_each_entry(c, &p->children, sibling) {
418 if (c->mm == p->mm)
419 continue;
420 if (!oom_kill_task(c))
421 return 0;
422 }
423 return oom_kill_task(p);
424 }
425
426 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
mem_cgroup_out_of_memory(struct mem_cgroup * mem,gfp_t gfp_mask)427 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
428 {
429 unsigned long points = 0;
430 struct task_struct *p;
431
432 read_lock(&tasklist_lock);
433 retry:
434 p = select_bad_process(&points, mem);
435 if (PTR_ERR(p) == -1UL)
436 goto out;
437
438 if (!p)
439 p = current;
440
441 if (oom_kill_process(p, gfp_mask, 0, points, mem,
442 "Memory cgroup out of memory"))
443 goto retry;
444 out:
445 read_unlock(&tasklist_lock);
446 }
447 #endif
448
449 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
450
register_oom_notifier(struct notifier_block * nb)451 int register_oom_notifier(struct notifier_block *nb)
452 {
453 return blocking_notifier_chain_register(&oom_notify_list, nb);
454 }
455 EXPORT_SYMBOL_GPL(register_oom_notifier);
456
unregister_oom_notifier(struct notifier_block * nb)457 int unregister_oom_notifier(struct notifier_block *nb)
458 {
459 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
460 }
461 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
462
463 /*
464 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
465 * if a parallel OOM killing is already taking place that includes a zone in
466 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
467 */
try_set_zone_oom(struct zonelist * zonelist,gfp_t gfp_mask)468 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
469 {
470 struct zoneref *z;
471 struct zone *zone;
472 int ret = 1;
473
474 spin_lock(&zone_scan_lock);
475 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
476 if (zone_is_oom_locked(zone)) {
477 ret = 0;
478 goto out;
479 }
480 }
481
482 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
483 /*
484 * Lock each zone in the zonelist under zone_scan_lock so a
485 * parallel invocation of try_set_zone_oom() doesn't succeed
486 * when it shouldn't.
487 */
488 zone_set_flag(zone, ZONE_OOM_LOCKED);
489 }
490
491 out:
492 spin_unlock(&zone_scan_lock);
493 return ret;
494 }
495
496 /*
497 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
498 * allocation attempts with zonelists containing them may now recall the OOM
499 * killer, if necessary.
500 */
clear_zonelist_oom(struct zonelist * zonelist,gfp_t gfp_mask)501 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
502 {
503 struct zoneref *z;
504 struct zone *zone;
505
506 spin_lock(&zone_scan_lock);
507 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
508 zone_clear_flag(zone, ZONE_OOM_LOCKED);
509 }
510 spin_unlock(&zone_scan_lock);
511 }
512
513 /*
514 * Must be called with tasklist_lock held for read.
515 */
__out_of_memory(gfp_t gfp_mask,int order)516 static void __out_of_memory(gfp_t gfp_mask, int order)
517 {
518 if (sysctl_oom_kill_allocating_task) {
519 oom_kill_process(current, gfp_mask, order, 0, NULL,
520 "Out of memory (oom_kill_allocating_task)");
521
522 } else {
523 unsigned long points;
524 struct task_struct *p;
525
526 retry:
527 /*
528 * Rambo mode: Shoot down a process and hope it solves whatever
529 * issues we may have.
530 */
531 p = select_bad_process(&points, NULL);
532
533 if (PTR_ERR(p) == -1UL)
534 return;
535
536 /* Found nothing?!?! Either we hang forever, or we panic. */
537 if (!p) {
538 read_unlock(&tasklist_lock);
539 panic("Out of memory and no killable processes...\n");
540 }
541
542 if (oom_kill_process(p, gfp_mask, order, points, NULL,
543 "Out of memory"))
544 goto retry;
545 }
546 }
547
548 /*
549 * pagefault handler calls into here because it is out of memory but
550 * doesn't know exactly how or why.
551 */
pagefault_out_of_memory(void)552 void pagefault_out_of_memory(void)
553 {
554 unsigned long freed = 0;
555
556 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
557 if (freed > 0)
558 /* Got some memory back in the last second. */
559 return;
560
561 /*
562 * If this is from memcg, oom-killer is already invoked.
563 * and not worth to go system-wide-oom.
564 */
565 if (mem_cgroup_oom_called(current))
566 goto rest_and_return;
567
568 if (sysctl_panic_on_oom)
569 panic("out of memory from page fault. panic_on_oom is selected.\n");
570
571 read_lock(&tasklist_lock);
572 __out_of_memory(0, 0); /* unknown gfp_mask and order */
573 read_unlock(&tasklist_lock);
574
575 /*
576 * Give "p" a good chance of killing itself before we
577 * retry to allocate memory.
578 */
579 rest_and_return:
580 if (!test_thread_flag(TIF_MEMDIE))
581 schedule_timeout_uninterruptible(1);
582 }
583
584 /**
585 * out_of_memory - kill the "best" process when we run out of memory
586 * @zonelist: zonelist pointer
587 * @gfp_mask: memory allocation flags
588 * @order: amount of memory being requested as a power of 2
589 *
590 * If we run out of memory, we have the choice between either
591 * killing a random task (bad), letting the system crash (worse)
592 * OR try to be smart about which process to kill. Note that we
593 * don't have to be perfect here, we just have to be good.
594 */
out_of_memory(struct zonelist * zonelist,gfp_t gfp_mask,int order)595 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
596 {
597 unsigned long freed = 0;
598 enum oom_constraint constraint;
599
600 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
601 if (freed > 0)
602 /* Got some memory back in the last second. */
603 return;
604
605 if (sysctl_panic_on_oom == 2)
606 panic("out of memory. Compulsory panic_on_oom is selected.\n");
607
608 /*
609 * Check if there were limitations on the allocation (only relevant for
610 * NUMA) that may require different handling.
611 */
612 constraint = constrained_alloc(zonelist, gfp_mask);
613 read_lock(&tasklist_lock);
614
615 switch (constraint) {
616 case CONSTRAINT_MEMORY_POLICY:
617 oom_kill_process(current, gfp_mask, order, 0, NULL,
618 "No available memory (MPOL_BIND)");
619 break;
620
621 case CONSTRAINT_NONE:
622 if (sysctl_panic_on_oom)
623 panic("out of memory. panic_on_oom is selected\n");
624 /* Fall-through */
625 case CONSTRAINT_CPUSET:
626 __out_of_memory(gfp_mask, order);
627 break;
628 }
629
630 read_unlock(&tasklist_lock);
631
632 /*
633 * Give "p" a good chance of killing itself before we
634 * retry to allocate memory unless "p" is current
635 */
636 if (!test_thread_flag(TIF_MEMDIE))
637 schedule_timeout_uninterruptible(1);
638 }
639