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>
30 int sysctl_panic_on_oom;
31 int sysctl_oom_kill_allocating_task;
32 int sysctl_oom_dump_tasks;
33 static DEFINE_SPINLOCK(zone_scan_mutex);
37 * badness - calculate a numeric value for how bad this task has been
38 * @p: task struct of which task we should calculate
39 * @uptime: current uptime in seconds
41 * The formula used is relatively simple and documented inline in the
42 * function. The main rationale is that we want to select a good task
43 * to kill when we run out of memory.
45 * Good in this context means that:
46 * 1) we lose the minimum amount of work done
47 * 2) we recover a large amount of memory
48 * 3) we don't kill anything innocent of eating tons of memory
49 * 4) we want to kill the minimum amount of processes (one)
50 * 5) we try to kill the process the user expects us to kill, this
51 * algorithm has been meticulously tuned to meet the principle
52 * of least surprise ... (be careful when you change it)
55 unsigned long badness(struct task_struct *p, unsigned long uptime,
56 struct mem_cgroup *mem)
58 unsigned long points, cpu_time, run_time, s;
60 struct task_struct *child;
70 * The memory size of the process is the basis for the badness.
72 points = mm->total_vm;
75 * After this unlock we can no longer dereference local variable `mm'
80 * swapoff can easily use up all memory, so kill those first.
82 if (p->flags & PF_SWAPOFF)
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.
93 list_for_each_entry(child, &p->children, sibling) {
95 if (child->mm != mm && child->mm)
96 points += child->mm->total_vm/2 + 1;
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.
105 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
108 if (uptime >= p->start_time.tv_sec)
109 run_time = (uptime - p->start_time.tv_sec) >> 10;
113 s = int_sqrt(cpu_time);
116 s = 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 (__capable(p, CAP_SYS_ADMIN) || __capable(p, CAP_SYS_RESOURCE))
135 * We don't want to kill a process with direct hardware access.
136 * Not only could that mess up the hardware, but usually users
137 * tend to only have this flag set on applications they think
140 if (__capable(p, CAP_SYS_RAWIO))
144 * If p's nodes don't overlap ours, it may still help to kill p
145 * because p may have allocated or otherwise mapped memory on
146 * this node before. However it will be less likely.
148 if (!cpuset_mems_allowed_intersects(current, p))
152 * Adjust the score by oomkilladj.
155 if (p->oomkilladj > 0) {
158 points <<= p->oomkilladj;
160 points >>= -(p->oomkilladj);
164 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
165 p->pid, p->comm, points);
171 * Determine the type of allocation constraint.
173 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
178 nodemask_t nodes = node_states[N_HIGH_MEMORY];
180 for (z = zonelist->zones; *z; z++)
181 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
182 node_clear(zone_to_nid(*z), nodes);
184 return CONSTRAINT_CPUSET;
186 if (!nodes_empty(nodes))
187 return CONSTRAINT_MEMORY_POLICY;
190 return CONSTRAINT_NONE;
194 * Simple selection loop. We chose the process with the highest
195 * number of 'points'. We expect the caller will lock the tasklist.
197 * (not docbooked, we don't want this one cluttering up the manual)
199 static struct task_struct *select_bad_process(unsigned long *ppoints,
200 struct mem_cgroup *mem)
202 struct task_struct *g, *p;
203 struct task_struct *chosen = NULL;
204 struct timespec uptime;
207 do_posix_clock_monotonic_gettime(&uptime);
208 do_each_thread(g, p) {
209 unsigned long points;
212 * skip kernel threads and tasks which have already released
217 /* skip the init task */
218 if (is_global_init(p))
220 if (mem && !task_in_mem_cgroup(p, mem))
224 * This task already has access to memory reserves and is
225 * being killed. Don't allow any other task access to the
228 * Note: this may have a chance of deadlock if it gets
229 * blocked waiting for another task which itself is waiting
230 * for memory. Is there a better alternative?
232 if (test_tsk_thread_flag(p, TIF_MEMDIE))
233 return ERR_PTR(-1UL);
236 * This is in the process of releasing memory so wait for it
237 * to finish before killing some other task by mistake.
239 * However, if p is the current task, we allow the 'kill' to
240 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
241 * which will allow it to gain access to memory reserves in
242 * the process of exiting and releasing its resources.
243 * Otherwise we could get an easy OOM deadlock.
245 if (p->flags & PF_EXITING) {
247 return ERR_PTR(-1UL);
250 *ppoints = ULONG_MAX;
253 if (p->oomkilladj == OOM_DISABLE)
256 points = badness(p, uptime.tv_sec, mem);
257 if (points > *ppoints || !chosen) {
261 } while_each_thread(g, p);
267 * Dumps the current memory state of all system tasks, excluding kernel threads.
268 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
271 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
274 * Call with tasklist_lock read-locked.
276 static void dump_tasks(const struct mem_cgroup *mem)
278 struct task_struct *g, *p;
280 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
282 do_each_thread(g, p) {
284 * total_vm and rss sizes do not exist for tasks with a
285 * detached mm so there's no need to report them.
289 if (mem && !task_in_mem_cgroup(p, mem))
293 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
294 p->pid, p->uid, p->tgid, p->mm->total_vm,
295 get_mm_rss(p->mm), (int)task_cpu(p), p->oomkilladj,
298 } while_each_thread(g, p);
302 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
303 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
306 static void __oom_kill_task(struct task_struct *p, int verbose)
308 if (is_global_init(p)) {
310 printk(KERN_WARNING "tried to kill init!\n");
316 printk(KERN_WARNING "tried to kill an mm-less task!\n");
321 printk(KERN_ERR "Killed process %d (%s)\n",
322 task_pid_nr(p), p->comm);
325 * We give our sacrificial lamb high priority and access to
326 * all the memory it needs. That way it should be able to
327 * exit() and clear out its resources quickly...
329 p->rt.time_slice = HZ;
330 set_tsk_thread_flag(p, TIF_MEMDIE);
332 force_sig(SIGKILL, p);
335 static int oom_kill_task(struct task_struct *p)
337 struct mm_struct *mm;
338 struct task_struct *g, *q;
342 /* WARNING: mm may not be dereferenced since we did not obtain its
343 * value from get_task_mm(p). This is OK since all we need to do is
344 * compare mm to q->mm below.
346 * Furthermore, even if mm contains a non-NULL value, p->mm may
347 * change to NULL at any time since we do not hold task_lock(p).
348 * However, this is of no concern to us.
355 * Don't kill the process if any threads are set to OOM_DISABLE
357 do_each_thread(g, q) {
358 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
360 } while_each_thread(g, q);
362 __oom_kill_task(p, 1);
365 * kill all processes that share the ->mm (i.e. all threads),
366 * but are in a different thread group. Don't let them have access
367 * to memory reserves though, otherwise we might deplete all memory.
369 do_each_thread(g, q) {
370 if (q->mm == mm && !same_thread_group(q, p))
371 force_sig(SIGKILL, q);
372 } while_each_thread(g, q);
377 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
378 unsigned long points, struct mem_cgroup *mem,
381 struct task_struct *c;
383 if (printk_ratelimit()) {
384 printk(KERN_WARNING "%s invoked oom-killer: "
385 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
386 current->comm, gfp_mask, order, current->oomkilladj);
389 if (sysctl_oom_dump_tasks)
394 * If the task is already exiting, don't alarm the sysadmin or kill
395 * its children or threads, just set TIF_MEMDIE so it can die quickly
397 if (p->flags & PF_EXITING) {
398 __oom_kill_task(p, 0);
402 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
403 message, task_pid_nr(p), p->comm, points);
405 /* Try to kill a child first */
406 list_for_each_entry(c, &p->children, sibling) {
409 if (!oom_kill_task(c))
412 return oom_kill_task(p);
415 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
416 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
418 unsigned long points = 0;
419 struct task_struct *p;
424 p = select_bad_process(&points, mem);
425 if (PTR_ERR(p) == -1UL)
431 if (oom_kill_process(p, gfp_mask, 0, points, mem,
432 "Memory cgroup out of memory"))
440 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
442 int register_oom_notifier(struct notifier_block *nb)
444 return blocking_notifier_chain_register(&oom_notify_list, nb);
446 EXPORT_SYMBOL_GPL(register_oom_notifier);
448 int unregister_oom_notifier(struct notifier_block *nb)
450 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
452 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
455 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
456 * if a parallel OOM killing is already taking place that includes a zone in
457 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
459 int try_set_zone_oom(struct zonelist *zonelist)
466 spin_lock(&zone_scan_mutex);
468 if (zone_is_oom_locked(*z)) {
472 } while (*(++z) != NULL);
475 * Lock each zone in the zonelist under zone_scan_mutex so a parallel
476 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
480 zone_set_flag(*z, ZONE_OOM_LOCKED);
481 } while (*(++z) != NULL);
483 spin_unlock(&zone_scan_mutex);
488 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
489 * allocation attempts with zonelists containing them may now recall the OOM
490 * killer, if necessary.
492 void clear_zonelist_oom(struct zonelist *zonelist)
498 spin_lock(&zone_scan_mutex);
500 zone_clear_flag(*z, ZONE_OOM_LOCKED);
501 } while (*(++z) != NULL);
502 spin_unlock(&zone_scan_mutex);
506 * out_of_memory - kill the "best" process when we run out of memory
508 * If we run out of memory, we have the choice between either
509 * killing a random task (bad), letting the system crash (worse)
510 * OR try to be smart about which process to kill. Note that we
511 * don't have to be perfect here, we just have to be good.
513 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
515 struct task_struct *p;
516 unsigned long points = 0;
517 unsigned long freed = 0;
518 enum oom_constraint constraint;
520 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
522 /* Got some memory back in the last second. */
525 if (sysctl_panic_on_oom == 2)
526 panic("out of memory. Compulsory panic_on_oom is selected.\n");
529 * Check if there were limitations on the allocation (only relevant for
530 * NUMA) that may require different handling.
532 constraint = constrained_alloc(zonelist, gfp_mask);
533 read_lock(&tasklist_lock);
535 switch (constraint) {
536 case CONSTRAINT_MEMORY_POLICY:
537 oom_kill_process(current, gfp_mask, order, points, NULL,
538 "No available memory (MPOL_BIND)");
541 case CONSTRAINT_NONE:
542 if (sysctl_panic_on_oom)
543 panic("out of memory. panic_on_oom is selected\n");
545 case CONSTRAINT_CPUSET:
546 if (sysctl_oom_kill_allocating_task) {
547 oom_kill_process(current, gfp_mask, order, points, NULL,
548 "Out of memory (oom_kill_allocating_task)");
553 * Rambo mode: Shoot down a process and hope it solves whatever
554 * issues we may have.
556 p = select_bad_process(&points, NULL);
558 if (PTR_ERR(p) == -1UL)
561 /* Found nothing?!?! Either we hang forever, or we panic. */
563 read_unlock(&tasklist_lock);
564 panic("Out of memory and no killable processes...\n");
567 if (oom_kill_process(p, gfp_mask, order, points, NULL,
575 read_unlock(&tasklist_lock);
578 * Give "p" a good chance of killing itself before we
579 * retry to allocate memory unless "p" is current
581 if (!test_thread_flag(TIF_MEMDIE))
582 schedule_timeout_uninterruptible(1);