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.
19 #include <linux/sched.h>
20 #include <linux/swap.h>
21 #include <linux/timex.h>
22 #include <linux/jiffies.h>
23 #include <linux/cpuset.h>
25 int sysctl_panic_on_oom;
29 * badness - calculate a numeric value for how bad this task has been
30 * @p: task struct of which task we should calculate
31 * @uptime: current uptime in seconds
33 * The formula used is relatively simple and documented inline in the
34 * function. The main rationale is that we want to select a good task
35 * to kill when we run out of memory.
37 * Good in this context means that:
38 * 1) we lose the minimum amount of work done
39 * 2) we recover a large amount of memory
40 * 3) we don't kill anything innocent of eating tons of memory
41 * 4) we want to kill the minimum amount of processes (one)
42 * 5) we try to kill the process the user expects us to kill, this
43 * algorithm has been meticulously tuned to meet the principle
44 * of least surprise ... (be careful when you change it)
47 unsigned long badness(struct task_struct *p, unsigned long uptime)
49 unsigned long points, cpu_time, run_time, s;
51 struct task_struct *child;
61 * swapoff can easily use up all memory, so kill those first.
63 if (p->flags & PF_SWAPOFF)
67 * The memory size of the process is the basis for the badness.
69 points = mm->total_vm;
72 * After this unlock we can no longer dereference local variable `mm'
76 /* FIXME: add vserver badness ;) */
79 * Processes which fork a lot of child processes are likely
80 * a good choice. We add half the vmsize of the children if they
81 * have an own mm. This prevents forking servers to flood the
82 * machine with an endless amount of children. In case a single
83 * child is eating the vast majority of memory, adding only half
84 * to the parents will make the child our kill candidate of choice.
86 list_for_each_entry(child, &p->children, sibling) {
88 if (child->mm != mm && child->mm)
89 points += child->mm->total_vm/2 + 1;
94 * CPU time is in tens of seconds and run time is in thousands
95 * of seconds. There is no particular reason for this other than
96 * that it turned out to work very well in practice.
98 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
101 if (uptime >= p->start_time.tv_sec)
102 run_time = (uptime - p->start_time.tv_sec) >> 10;
106 s = int_sqrt(cpu_time);
109 s = int_sqrt(int_sqrt(run_time));
114 * Niced processes are most likely less important, so double
115 * their badness points.
117 if (task_nice(p) > 0)
121 * Superuser processes are usually more important, so we make it
122 * less likely that we kill those.
124 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
125 p->uid == 0 || p->euid == 0)
129 * We don't want to kill a process with direct hardware access.
130 * Not only could that mess up the hardware, but usually users
131 * tend to only have this flag set on applications they think
134 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
138 * If p's nodes don't overlap ours, it may still help to kill p
139 * because p may have allocated or otherwise mapped memory on
140 * this node before. However it will be less likely.
142 if (!cpuset_excl_nodes_overlap(p))
146 * Adjust the score by oomkilladj.
149 if (p->oomkilladj > 0)
150 points <<= p->oomkilladj;
152 points >>= -(p->oomkilladj);
156 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
157 p->pid, p->comm, points);
162 #if defined(CONFIG_OOM_PANIC) && defined(CONFIG_OOM_KILLER)
163 #warning Only define OOM_PANIC or OOM_KILLER; not both
166 #ifdef CONFIG_OOM_KILLER
168 * Types of limitations to the nodes from which allocations may occur
170 #define CONSTRAINT_NONE 1
171 #define CONSTRAINT_MEMORY_POLICY 2
172 #define CONSTRAINT_CPUSET 3
175 * Determine the type of allocation constraint.
177 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
181 nodemask_t nodes = node_online_map;
183 for (z = zonelist->zones; *z; z++)
184 if (cpuset_zone_allowed(*z, gfp_mask))
185 node_clear((*z)->zone_pgdat->node_id,
188 return CONSTRAINT_CPUSET;
190 if (!nodes_empty(nodes))
191 return CONSTRAINT_MEMORY_POLICY;
194 return CONSTRAINT_NONE;
198 * Simple selection loop. We chose the process with the highest
199 * number of 'points'. We expect the caller will lock the tasklist.
201 * (not docbooked, we don't want this one cluttering up the manual)
203 static struct task_struct *select_bad_process(unsigned long *ppoints)
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 */
219 /* skip the init task with pid == 1 */
223 * This is in the process of releasing memory so wait for it
224 * to finish before killing some other task by mistake.
226 * However, if p is the current task, we allow the 'kill' to
227 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
228 * which will allow it to gain access to memory reserves in
229 * the process of exiting and releasing its resources.
230 * Otherwise we could get an OOM deadlock.
232 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
233 p->flags & PF_EXITING;
235 /* PF_DEAD tasks have already released their mm */
236 if (p->flags & PF_DEAD)
238 if (p->flags & PF_EXITING && p == current) {
240 *ppoints = ULONG_MAX;
243 return ERR_PTR(-1UL);
245 if (p->oomkilladj == OOM_DISABLE)
248 points = badness(p, uptime.tv_sec);
249 if (points > *ppoints || !chosen) {
253 } while_each_thread(g, p);
258 * We must be careful though to never send SIGKILL a process with
259 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
260 * we select a process with CAP_SYS_RAW_IO set).
262 static void __oom_kill_task(struct task_struct *p, const char *message)
266 printk(KERN_WARNING "tried to kill init!\n");
271 if (!p->mm || p->mm == &init_mm) {
273 printk(KERN_WARNING "tried to kill an mm-less task!\n");
279 printk(KERN_ERR "%s: Killed process %d (%s).\n",
280 message, p->pid, p->comm);
283 * We give our sacrificial lamb high priority and access to
284 * all the memory it needs. That way it should be able to
285 * exit() and clear out its resources quickly...
288 set_tsk_thread_flag(p, TIF_MEMDIE);
290 force_sig(SIGKILL, p);
293 static int oom_kill_task(struct task_struct *p, const char *message)
295 struct mm_struct *mm;
296 struct task_struct *g, *q;
300 /* WARNING: mm may not be dereferenced since we did not obtain its
301 * value from get_task_mm(p). This is OK since all we need to do is
302 * compare mm to q->mm below.
304 * Furthermore, even if mm contains a non-NULL value, p->mm may
305 * change to NULL at any time since we do not hold task_lock(p).
306 * However, this is of no concern to us.
309 if (mm == NULL || mm == &init_mm)
312 __oom_kill_task(p, message);
314 * kill all processes that share the ->mm (i.e. all threads),
315 * but are in a different thread group
318 if (q->mm == mm && q->tgid != p->tgid)
319 __oom_kill_task(q, message);
320 while_each_thread(g, q);
325 static int oom_kill_process(struct task_struct *p, unsigned long points,
328 struct task_struct *c;
329 struct list_head *tsk;
332 * If the task is already exiting, don't alarm the sysadmin or kill
333 * its children or threads, just set TIF_MEMDIE so it can die quickly
335 if (p->flags & PF_EXITING) {
336 __oom_kill_task(p, NULL);
340 /* Try to kill a child first */
341 list_for_each(tsk, &p->children) {
342 c = list_entry(tsk, struct task_struct, sibling);
345 if (!oom_kill_task(c, message))
348 return oom_kill_task(p, message);
351 int should_oom_kill(void)
353 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
354 static unsigned long first, last, count, lastkill;
355 unsigned long now, since;
358 spin_lock(&oom_lock);
364 * If it's been a long time since last failure,
371 * If we haven't tried for at least one second,
372 * we're not really oom.
379 * If we have gotten only a few failures,
380 * we're not really oom.
386 * If we just killed a process, wait a while
387 * to give that task a chance to exit. This
388 * avoids killing multiple processes needlessly.
390 since = now - lastkill;
395 * Ok, really out of memory. Kill something.
402 * We dropped the lock above, so check to be sure the variable
403 * first only ever increases to prevent false OOM's.
405 if (time_after(now, first))
410 spin_unlock(&oom_lock);
415 * out_of_memory - kill the "best" process when we run out of memory
417 * If we run out of memory, we have the choice between either
418 * killing a random task (bad), letting the system crash (worse)
419 * OR try to be smart about which process to kill. Note that we
420 * don't have to be perfect here, we just have to be good.
422 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
424 struct task_struct *p;
425 unsigned long points = 0;
427 if (printk_ratelimit()) {
428 printk(KERN_WARNING "%s invoked oom-killer: "
429 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
430 current->comm, gfp_mask, order, current->oomkilladj);
435 if (!should_oom_kill())
439 read_lock(&tasklist_lock);
442 * Check if there were limitations on the allocation (only relevant for
443 * NUMA) that may require different handling.
445 switch (constrained_alloc(zonelist, gfp_mask)) {
446 case CONSTRAINT_MEMORY_POLICY:
447 oom_kill_process(current, points,
448 "No available memory (MPOL_BIND)");
451 case CONSTRAINT_CPUSET:
452 oom_kill_process(current, points,
453 "No available memory in cpuset");
456 case CONSTRAINT_NONE:
457 if (sysctl_panic_on_oom)
458 panic("out of memory. panic_on_oom is selected\n");
461 * Rambo mode: Shoot down a process and hope it solves whatever
462 * issues we may have.
464 p = select_bad_process(&points);
466 if (PTR_ERR(p) == -1UL)
469 /* Found nothing?!?! Either we hang forever, or we panic. */
471 read_unlock(&tasklist_lock);
473 panic("Out of memory and no killable processes...\n");
476 if (oom_kill_process(p, points, "Out of memory"))
483 read_unlock(&tasklist_lock);
487 * Give "p" a good chance of killing itself before we
488 * retry to allocate memory unless "p" is current
490 if (!test_thread_flag(TIF_MEMDIE))
491 schedule_timeout_uninterruptible(1);
493 #endif /* CONFIG_OOM_KILLER */
495 #ifdef CONFIG_OOM_PANIC
497 * out_of_memory - panic if the system out of memory?
499 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
502 * oom_lock protects out_of_memory()'s static variables.
503 * It's a global lock; this is not performance-critical.
505 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
506 static unsigned long count;
508 spin_lock(&oom_lock);
511 * If we have gotten only a few failures,
512 * we're not really oom.
516 * Ok, really out of memory. Panic.
519 printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
522 panic("Out Of Memory");
524 spin_unlock(&oom_lock);
526 #endif /* CONFIG_OOM_PANIC */