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...
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.
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.
18 #include <linux/oom.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>
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);
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
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.
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)
56 unsigned long badness(struct task_struct *p, unsigned long uptime)
58 unsigned long points, cpu_time, run_time;
60 struct task_struct *child;
69 oom_adj = mm->oom_adj;
72 * The memory size of the process is the basis for the badness.
74 points = mm->total_vm;
77 * After this unlock we can no longer dereference local variable `mm'
82 * swapoff can easily use up all memory, so kill those first.
84 if (p->flags & PF_SWAPOFF)
88 * Processes which fork a lot of child processes are likely
89 * a good choice. We add half the vmsize of the children if they
90 * have an own mm. This prevents forking servers to flood the
91 * machine with an endless amount of children. In case a single
92 * child is eating the vast majority of memory, adding only half
93 * to the parents will make the child our kill candidate of choice.
95 list_for_each_entry(child, &p->children, sibling) {
97 if (child->mm != mm && child->mm)
98 points += child->mm->total_vm/2 + 1;
103 * CPU time is in tens of seconds and run time is in thousands
104 * of seconds. There is no particular reason for this other than
105 * that it turned out to work very well in practice.
107 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
110 if (uptime >= p->start_time.tv_sec)
111 run_time = (uptime - p->start_time.tv_sec) >> 10;
116 points /= int_sqrt(cpu_time);
118 points /= int_sqrt(int_sqrt(run_time));
121 * Niced processes are most likely less important, so double
122 * their badness points.
124 if (task_nice(p) > 0)
128 * Superuser processes are usually more important, so we make it
129 * less likely that we kill those.
131 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
132 has_capability_noaudit(p, CAP_SYS_RESOURCE))
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
141 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
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.
149 if (!cpuset_mems_allowed_intersects(current, p))
153 * Adjust the score by oom_adj.
161 points >>= -(oom_adj);
165 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
166 p->pid, p->comm, points);
172 * Determine the type of allocation constraint.
174 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
180 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
181 nodemask_t nodes = node_states[N_HIGH_MEMORY];
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);
187 return CONSTRAINT_CPUSET;
189 if (!nodes_empty(nodes))
190 return CONSTRAINT_MEMORY_POLICY;
193 return CONSTRAINT_NONE;
197 * Simple selection loop. We chose the process with the highest
198 * number of 'points'. We expect the caller will lock the tasklist.
200 * (not docbooked, we don't want this one cluttering up the manual)
202 static struct task_struct *select_bad_process(unsigned long *ppoints,
203 struct mem_cgroup *mem)
205 struct task_struct *g, *p;
206 struct task_struct *chosen = NULL;
207 struct timespec uptime;
210 do_posix_clock_monotonic_gettime(&uptime);
211 do_each_thread(g, p) {
212 unsigned long points;
215 * skip kernel threads and tasks which have already released
220 /* skip the init task */
221 if (is_global_init(p))
223 if (mem && !task_in_mem_cgroup(p, mem))
227 * This task already has access to memory reserves and is
228 * being killed. Don't allow any other task access to the
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?
235 if (test_tsk_thread_flag(p, TIF_MEMDIE))
236 return ERR_PTR(-1UL);
239 * This is in the process of releasing memory so wait for it
240 * to finish before killing some other task by mistake.
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.
248 if (p->flags & PF_EXITING) {
250 return ERR_PTR(-1UL);
253 *ppoints = ULONG_MAX;
257 if (p->mm && p->mm->oom_adj == OOM_DISABLE) {
263 points = badness(p, uptime.tv_sec);
264 if (points > *ppoints || !chosen) {
268 } while_each_thread(g, p);
274 * dump_tasks - dump current memory state of all system tasks
275 * @mem: target memory controller
277 * Dumps the current memory state of all system tasks, excluding kernel threads.
278 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
281 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
284 * Call with tasklist_lock read-locked.
286 static void dump_tasks(const struct mem_cgroup *mem)
288 struct task_struct *g, *p;
290 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
292 do_each_thread(g, p) {
293 struct mm_struct *mm;
295 if (mem && !task_in_mem_cgroup(p, mem))
297 if (!thread_group_leader(p))
304 * total_vm and rss sizes do not exist for tasks with no
305 * mm so there's no need to report them; they can't be
311 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
312 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
313 get_mm_rss(mm), (int)task_cpu(p), mm->oom_adj, p->comm);
315 } while_each_thread(g, p);
319 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
320 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
323 static void __oom_kill_task(struct task_struct *p, int verbose)
325 if (is_global_init(p)) {
327 printk(KERN_WARNING "tried to kill init!\n");
333 printk(KERN_WARNING "tried to kill an mm-less task!\n");
338 printk(KERN_ERR "Killed process %d (%s)\n",
339 task_pid_nr(p), p->comm);
342 * We give our sacrificial lamb high priority and access to
343 * all the memory it needs. That way it should be able to
344 * exit() and clear out its resources quickly...
346 p->rt.time_slice = HZ;
347 set_tsk_thread_flag(p, TIF_MEMDIE);
349 force_sig(SIGKILL, p);
352 static int oom_kill_task(struct task_struct *p)
354 struct mm_struct *mm;
355 struct task_struct *g, *q;
359 /* WARNING: mm may not be dereferenced since we did not obtain its
360 * value from get_task_mm(p). This is OK since all we need to do is
361 * compare mm to q->mm below.
363 * Furthermore, even if mm contains a non-NULL value, p->mm may
364 * change to NULL at any time since we do not hold task_lock(p).
365 * However, this is of no concern to us.
372 * Don't kill the process if any threads are set to OOM_DISABLE
374 do_each_thread(g, q) {
376 if (q->mm == mm && q->mm && q->mm->oom_adj == OOM_DISABLE) {
381 } while_each_thread(g, q);
383 __oom_kill_task(p, 1);
386 * kill all processes that share the ->mm (i.e. all threads),
387 * but are in a different thread group. Don't let them have access
388 * to memory reserves though, otherwise we might deplete all memory.
390 do_each_thread(g, q) {
391 if (q->mm == mm && !same_thread_group(q, p))
392 force_sig(SIGKILL, q);
393 } while_each_thread(g, q);
398 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
399 unsigned long points, struct mem_cgroup *mem,
402 struct task_struct *c;
404 if (printk_ratelimit()) {
406 printk(KERN_WARNING "%s invoked oom-killer: "
407 "gfp_mask=0x%x, order=%d, oom_adj=%d\n",
408 current->comm, gfp_mask, order,
409 current->mm ? current->mm->oom_adj : OOM_DISABLE);
410 cpuset_print_task_mems_allowed(current);
411 task_unlock(current);
413 mem_cgroup_print_oom_info(mem, current);
415 if (sysctl_oom_dump_tasks)
420 * If the task is already exiting, don't alarm the sysadmin or kill
421 * its children or threads, just set TIF_MEMDIE so it can die quickly
423 if (p->flags & PF_EXITING) {
424 __oom_kill_task(p, 0);
428 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
429 message, task_pid_nr(p), p->comm, points);
431 /* Try to kill a child first */
432 list_for_each_entry(c, &p->children, sibling) {
435 if (!oom_kill_task(c))
438 return oom_kill_task(p);
441 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
442 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
444 unsigned long points = 0;
445 struct task_struct *p;
447 read_lock(&tasklist_lock);
449 p = select_bad_process(&points, mem);
450 if (PTR_ERR(p) == -1UL)
456 if (oom_kill_process(p, gfp_mask, 0, points, mem,
457 "Memory cgroup out of memory"))
460 read_unlock(&tasklist_lock);
464 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
466 int register_oom_notifier(struct notifier_block *nb)
468 return blocking_notifier_chain_register(&oom_notify_list, nb);
470 EXPORT_SYMBOL_GPL(register_oom_notifier);
472 int unregister_oom_notifier(struct notifier_block *nb)
474 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
476 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
479 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
480 * if a parallel OOM killing is already taking place that includes a zone in
481 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
483 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
489 spin_lock(&zone_scan_lock);
490 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
491 if (zone_is_oom_locked(zone)) {
497 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
499 * Lock each zone in the zonelist under zone_scan_lock so a
500 * parallel invocation of try_set_zone_oom() doesn't succeed
503 zone_set_flag(zone, ZONE_OOM_LOCKED);
507 spin_unlock(&zone_scan_lock);
512 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
513 * allocation attempts with zonelists containing them may now recall the OOM
514 * killer, if necessary.
516 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
521 spin_lock(&zone_scan_lock);
522 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
523 zone_clear_flag(zone, ZONE_OOM_LOCKED);
525 spin_unlock(&zone_scan_lock);
529 * Must be called with tasklist_lock held for read.
531 static void __out_of_memory(gfp_t gfp_mask, int order)
533 struct task_struct *p;
534 unsigned long points;
536 if (sysctl_oom_kill_allocating_task)
537 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
538 "Out of memory (oom_kill_allocating_task)"))
542 * Rambo mode: Shoot down a process and hope it solves whatever
543 * issues we may have.
545 p = select_bad_process(&points, NULL);
547 if (PTR_ERR(p) == -1UL)
550 /* Found nothing?!?! Either we hang forever, or we panic. */
552 read_unlock(&tasklist_lock);
553 panic("Out of memory and no killable processes...\n");
556 if (oom_kill_process(p, gfp_mask, order, points, NULL,
562 * pagefault handler calls into here because it is out of memory but
563 * doesn't know exactly how or why.
565 void pagefault_out_of_memory(void)
567 unsigned long freed = 0;
569 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
571 /* Got some memory back in the last second. */
575 * If this is from memcg, oom-killer is already invoked.
576 * and not worth to go system-wide-oom.
578 if (mem_cgroup_oom_called(current))
579 goto rest_and_return;
581 if (sysctl_panic_on_oom)
582 panic("out of memory from page fault. panic_on_oom is selected.\n");
584 read_lock(&tasklist_lock);
585 __out_of_memory(0, 0); /* unknown gfp_mask and order */
586 read_unlock(&tasklist_lock);
589 * Give "p" a good chance of killing itself before we
590 * retry to allocate memory.
593 if (!test_thread_flag(TIF_MEMDIE))
594 schedule_timeout_uninterruptible(1);
598 * out_of_memory - kill the "best" process when we run out of memory
599 * @zonelist: zonelist pointer
600 * @gfp_mask: memory allocation flags
601 * @order: amount of memory being requested as a power of 2
603 * If we run out of memory, we have the choice between either
604 * killing a random task (bad), letting the system crash (worse)
605 * OR try to be smart about which process to kill. Note that we
606 * don't have to be perfect here, we just have to be good.
608 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
610 unsigned long freed = 0;
611 enum oom_constraint constraint;
613 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
615 /* Got some memory back in the last second. */
618 if (sysctl_panic_on_oom == 2)
619 panic("out of memory. Compulsory panic_on_oom is selected.\n");
622 * Check if there were limitations on the allocation (only relevant for
623 * NUMA) that may require different handling.
625 constraint = constrained_alloc(zonelist, gfp_mask);
626 read_lock(&tasklist_lock);
628 switch (constraint) {
629 case CONSTRAINT_MEMORY_POLICY:
630 oom_kill_process(current, gfp_mask, order, 0, NULL,
631 "No available memory (MPOL_BIND)");
634 case CONSTRAINT_NONE:
635 if (sysctl_panic_on_oom)
636 panic("out of memory. panic_on_oom is selected\n");
638 case CONSTRAINT_CPUSET:
639 __out_of_memory(gfp_mask, order);
643 read_unlock(&tasklist_lock);
646 * Give "p" a good chance of killing itself before we
647 * retry to allocate memory unless "p" is current
649 if (!test_thread_flag(TIF_MEMDIE))
650 schedule_timeout_uninterruptible(1);