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 * Is all threads of the target process nodes overlap ours?
40 static int has_intersects_mems_allowed(struct task_struct *tsk)
42 struct task_struct *t;
46 if (cpuset_mems_allowed_intersects(current, t))
55 * badness - calculate a numeric value for how bad this task has been
56 * @p: task struct of which task we should calculate
57 * @uptime: current uptime in seconds
59 * The formula used is relatively simple and documented inline in the
60 * function. The main rationale is that we want to select a good task
61 * to kill when we run out of memory.
63 * Good in this context means that:
64 * 1) we lose the minimum amount of work done
65 * 2) we recover a large amount of memory
66 * 3) we don't kill anything innocent of eating tons of memory
67 * 4) we want to kill the minimum amount of processes (one)
68 * 5) we try to kill the process the user expects us to kill, this
69 * algorithm has been meticulously tuned to meet the principle
70 * of least surprise ... (be careful when you change it)
73 unsigned long badness(struct task_struct *p, unsigned long uptime)
75 unsigned long points, cpu_time, run_time;
77 struct task_struct *child;
78 int oom_adj = p->signal->oom_adj;
79 struct task_cputime task_time;
83 if (oom_adj == OOM_DISABLE)
94 * The memory size of the process is the basis for the badness.
96 points = mm->total_vm;
99 * After this unlock we can no longer dereference local variable `mm'
104 * swapoff can easily use up all memory, so kill those first.
106 if (p->flags & PF_OOM_ORIGIN)
110 * Processes which fork a lot of child processes are likely
111 * a good choice. We add half the vmsize of the children if they
112 * have an own mm. This prevents forking servers to flood the
113 * machine with an endless amount of children. In case a single
114 * child is eating the vast majority of memory, adding only half
115 * to the parents will make the child our kill candidate of choice.
117 list_for_each_entry(child, &p->children, sibling) {
119 if (child->mm != mm && child->mm)
120 points += child->mm->total_vm/2 + 1;
125 * CPU time is in tens of seconds and run time is in thousands
126 * of seconds. There is no particular reason for this other than
127 * that it turned out to work very well in practice.
129 thread_group_cputime(p, &task_time);
130 utime = cputime_to_jiffies(task_time.utime);
131 stime = cputime_to_jiffies(task_time.stime);
132 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
135 if (uptime >= p->start_time.tv_sec)
136 run_time = (uptime - p->start_time.tv_sec) >> 10;
141 points /= int_sqrt(cpu_time);
143 points /= int_sqrt(int_sqrt(run_time));
146 * Niced processes are most likely less important, so double
147 * their badness points.
149 if (task_nice(p) > 0)
153 * Superuser processes are usually more important, so we make it
154 * less likely that we kill those.
156 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
157 has_capability_noaudit(p, CAP_SYS_RESOURCE))
161 * We don't want to kill a process with direct hardware access.
162 * Not only could that mess up the hardware, but usually users
163 * tend to only have this flag set on applications they think
166 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
170 * If p's nodes don't overlap ours, it may still help to kill p
171 * because p may have allocated or otherwise mapped memory on
172 * this node before. However it will be less likely.
174 if (!has_intersects_mems_allowed(p))
178 * Adjust the score by oom_adj.
186 points >>= -(oom_adj);
190 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
191 p->pid, p->comm, points);
197 * Determine the type of allocation constraint.
199 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
205 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
206 nodemask_t nodes = node_states[N_HIGH_MEMORY];
208 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
209 if (cpuset_zone_allowed_softwall(zone, gfp_mask))
210 node_clear(zone_to_nid(zone), nodes);
212 return CONSTRAINT_CPUSET;
214 if (!nodes_empty(nodes))
215 return CONSTRAINT_MEMORY_POLICY;
218 return CONSTRAINT_NONE;
222 * Simple selection loop. We chose the process with the highest
223 * number of 'points'. We expect the caller will lock the tasklist.
225 * (not docbooked, we don't want this one cluttering up the manual)
227 static struct task_struct *select_bad_process(unsigned long *ppoints,
228 struct mem_cgroup *mem)
230 struct task_struct *p;
231 struct task_struct *chosen = NULL;
232 struct timespec uptime;
235 do_posix_clock_monotonic_gettime(&uptime);
236 for_each_process(p) {
237 unsigned long points;
240 * skip kernel threads and tasks which have already released
245 /* skip the init task */
246 if (is_global_init(p))
248 if (mem && !task_in_mem_cgroup(p, mem))
252 * This task already has access to memory reserves and is
253 * being killed. Don't allow any other task access to the
256 * Note: this may have a chance of deadlock if it gets
257 * blocked waiting for another task which itself is waiting
258 * for memory. Is there a better alternative?
260 if (test_tsk_thread_flag(p, TIF_MEMDIE))
261 return ERR_PTR(-1UL);
264 * This is in the process of releasing memory so wait for it
265 * to finish before killing some other task by mistake.
267 * However, if p is the current task, we allow the 'kill' to
268 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
269 * which will allow it to gain access to memory reserves in
270 * the process of exiting and releasing its resources.
271 * Otherwise we could get an easy OOM deadlock.
273 if (p->flags & PF_EXITING) {
275 return ERR_PTR(-1UL);
278 *ppoints = ULONG_MAX;
281 if (p->signal->oom_adj == OOM_DISABLE)
284 points = badness(p, uptime.tv_sec);
285 if (points > *ppoints || !chosen) {
295 * dump_tasks - dump current memory state of all system tasks
296 * @mem: target memory controller
298 * Dumps the current memory state of all system tasks, excluding kernel threads.
299 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
302 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
305 * Call with tasklist_lock read-locked.
307 static void dump_tasks(const struct mem_cgroup *mem)
309 struct task_struct *g, *p;
311 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
313 do_each_thread(g, p) {
314 struct mm_struct *mm;
316 if (mem && !task_in_mem_cgroup(p, mem))
318 if (!thread_group_leader(p))
325 * total_vm and rss sizes do not exist for tasks with no
326 * mm so there's no need to report them; they can't be
332 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
333 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
334 get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj,
337 } while_each_thread(g, p);
341 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
342 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
345 static void __oom_kill_task(struct task_struct *p, int verbose)
347 if (is_global_init(p)) {
349 printk(KERN_WARNING "tried to kill init!\n");
355 printk(KERN_WARNING "tried to kill an mm-less task!\n");
360 printk(KERN_ERR "Killed process %d (%s)\n",
361 task_pid_nr(p), p->comm);
364 * We give our sacrificial lamb high priority and access to
365 * all the memory it needs. That way it should be able to
366 * exit() and clear out its resources quickly...
368 p->rt.time_slice = HZ;
369 set_tsk_thread_flag(p, TIF_MEMDIE);
371 force_sig(SIGKILL, p);
374 static int oom_kill_task(struct task_struct *p)
376 /* WARNING: mm may not be dereferenced since we did not obtain its
377 * value from get_task_mm(p). This is OK since all we need to do is
378 * compare mm to q->mm below.
380 * Furthermore, even if mm contains a non-NULL value, p->mm may
381 * change to NULL at any time since we do not hold task_lock(p).
382 * However, this is of no concern to us.
384 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
387 __oom_kill_task(p, 1);
392 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
393 unsigned long points, struct mem_cgroup *mem,
396 struct task_struct *c;
398 if (printk_ratelimit()) {
399 printk(KERN_WARNING "%s invoked oom-killer: "
400 "gfp_mask=0x%x, order=%d, oom_adj=%d\n",
401 current->comm, gfp_mask, order,
402 current->signal->oom_adj);
404 cpuset_print_task_mems_allowed(current);
405 task_unlock(current);
407 mem_cgroup_print_oom_info(mem, current);
409 if (sysctl_oom_dump_tasks)
414 * If the task is already exiting, don't alarm the sysadmin or kill
415 * its children or threads, just set TIF_MEMDIE so it can die quickly
417 if (p->flags & PF_EXITING) {
418 __oom_kill_task(p, 0);
422 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
423 message, task_pid_nr(p), p->comm, points);
425 /* Try to kill a child first */
426 list_for_each_entry(c, &p->children, sibling) {
429 if (!oom_kill_task(c))
432 return oom_kill_task(p);
435 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
436 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
438 unsigned long points = 0;
439 struct task_struct *p;
441 read_lock(&tasklist_lock);
443 p = select_bad_process(&points, mem);
444 if (PTR_ERR(p) == -1UL)
450 if (oom_kill_process(p, gfp_mask, 0, points, mem,
451 "Memory cgroup out of memory"))
454 read_unlock(&tasklist_lock);
458 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
460 int register_oom_notifier(struct notifier_block *nb)
462 return blocking_notifier_chain_register(&oom_notify_list, nb);
464 EXPORT_SYMBOL_GPL(register_oom_notifier);
466 int unregister_oom_notifier(struct notifier_block *nb)
468 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
470 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
473 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
474 * if a parallel OOM killing is already taking place that includes a zone in
475 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
477 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
483 spin_lock(&zone_scan_lock);
484 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
485 if (zone_is_oom_locked(zone)) {
491 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
493 * Lock each zone in the zonelist under zone_scan_lock so a
494 * parallel invocation of try_set_zone_oom() doesn't succeed
497 zone_set_flag(zone, ZONE_OOM_LOCKED);
501 spin_unlock(&zone_scan_lock);
506 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
507 * allocation attempts with zonelists containing them may now recall the OOM
508 * killer, if necessary.
510 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
515 spin_lock(&zone_scan_lock);
516 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
517 zone_clear_flag(zone, ZONE_OOM_LOCKED);
519 spin_unlock(&zone_scan_lock);
523 * Must be called with tasklist_lock held for read.
525 static void __out_of_memory(gfp_t gfp_mask, int order)
527 struct task_struct *p;
528 unsigned long points;
530 if (sysctl_oom_kill_allocating_task)
531 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
532 "Out of memory (oom_kill_allocating_task)"))
536 * Rambo mode: Shoot down a process and hope it solves whatever
537 * issues we may have.
539 p = select_bad_process(&points, NULL);
541 if (PTR_ERR(p) == -1UL)
544 /* Found nothing?!?! Either we hang forever, or we panic. */
546 read_unlock(&tasklist_lock);
547 panic("Out of memory and no killable processes...\n");
550 if (oom_kill_process(p, gfp_mask, order, points, NULL,
556 * pagefault handler calls into here because it is out of memory but
557 * doesn't know exactly how or why.
559 void pagefault_out_of_memory(void)
561 unsigned long freed = 0;
563 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
565 /* Got some memory back in the last second. */
569 * If this is from memcg, oom-killer is already invoked.
570 * and not worth to go system-wide-oom.
572 if (mem_cgroup_oom_called(current))
573 goto rest_and_return;
575 if (sysctl_panic_on_oom)
576 panic("out of memory from page fault. panic_on_oom is selected.\n");
578 read_lock(&tasklist_lock);
579 __out_of_memory(0, 0); /* unknown gfp_mask and order */
580 read_unlock(&tasklist_lock);
583 * Give "p" a good chance of killing itself before we
584 * retry to allocate memory.
587 if (!test_thread_flag(TIF_MEMDIE))
588 schedule_timeout_uninterruptible(1);
592 * out_of_memory - kill the "best" process when we run out of memory
593 * @zonelist: zonelist pointer
594 * @gfp_mask: memory allocation flags
595 * @order: amount of memory being requested as a power of 2
597 * If we run out of memory, we have the choice between either
598 * killing a random task (bad), letting the system crash (worse)
599 * OR try to be smart about which process to kill. Note that we
600 * don't have to be perfect here, we just have to be good.
602 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
604 unsigned long freed = 0;
605 enum oom_constraint constraint;
607 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
609 /* Got some memory back in the last second. */
612 if (sysctl_panic_on_oom == 2)
613 panic("out of memory. Compulsory panic_on_oom is selected.\n");
616 * Check if there were limitations on the allocation (only relevant for
617 * NUMA) that may require different handling.
619 constraint = constrained_alloc(zonelist, gfp_mask);
620 read_lock(&tasklist_lock);
622 switch (constraint) {
623 case CONSTRAINT_MEMORY_POLICY:
624 oom_kill_process(current, gfp_mask, order, 0, NULL,
625 "No available memory (MPOL_BIND)");
628 case CONSTRAINT_NONE:
629 if (sysctl_panic_on_oom)
630 panic("out of memory. panic_on_oom is selected\n");
632 case CONSTRAINT_CPUSET:
633 __out_of_memory(gfp_mask, order);
637 read_unlock(&tasklist_lock);
640 * Give "p" a good chance of killing itself before we
641 * retry to allocate memory unless "p" is current
643 if (!test_thread_flag(TIF_MEMDIE))
644 schedule_timeout_uninterruptible(1);