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 kswapd()
10 * in linux/mm/vmscan.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>
27 * oom_badness - calculate a numeric value for how bad this task has been
28 * @p: task struct of which task we should calculate
30 * The formula used is relatively simple and documented inline in the
31 * function. The main rationale is that we want to select a good task
32 * to kill when we run out of memory.
34 * Good in this context means that:
35 * 1) we lose the minimum amount of work done
36 * 2) we recover a large amount of memory
37 * 3) we don't kill anything innocent of eating tons of memory
38 * 4) we want to kill the minimum amount of processes (one)
39 * 5) we try to kill the process the user expects us to kill, this
40 * algorithm has been meticulously tuned to meet the principle
41 * of least surprise ... (be careful when you change it)
44 static int badness(struct task_struct *p)
46 int points, cpu_time, run_time, s;
51 if (p->flags & PF_MEMDIE)
54 * The memory size of the process is the basis for the badness.
56 points = p->mm->total_vm;
57 /* add vserver badness ;) */
60 * CPU time is in seconds and run time is in minutes. There is no
61 * particular reason for this other than that it turned out to work
62 * very well in practice.
64 cpu_time = (p->utime + p->stime) >> (SHIFT_HZ + 3);
65 run_time = (get_jiffies_64() - p->start_time) >> (SHIFT_HZ + 10);
67 s = int_sqrt(cpu_time);
70 s = int_sqrt(int_sqrt(run_time));
75 * Niced processes are most likely less important, so double
76 * their badness points.
82 * Superuser processes are usually more important, so we make it
83 * less likely that we kill those.
85 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
86 p->uid == 0 || p->euid == 0)
90 * We don't want to kill a process with direct hardware access.
91 * Not only could that mess up the hardware, but usually users
92 * tend to only have this flag set on applications they think
95 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
98 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
99 p->pid, p->comm, points);
105 * Simple selection loop. We chose the process with the highest
106 * number of 'points'. We expect the caller will lock the tasklist.
108 * (not docbooked, we don't want this one cluttering up the manual)
110 static struct task_struct * select_bad_process(void)
113 struct task_struct *g, *p;
114 struct task_struct *chosen = NULL;
118 int points = badness(p);
119 if (points > maxpoints) {
123 if (p->flags & PF_SWAPOFF)
126 while_each_thread(g, p);
131 * We must be careful though to never send SIGKILL a process with
132 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
133 * we select a process with CAP_SYS_RAW_IO set).
135 static void __oom_kill_task(task_t *p)
138 if (!p->mm || p->mm == &init_mm) {
140 printk(KERN_WARNING "tried to kill an mm-less task!\n");
145 printk(KERN_ERR "Out of Memory: Killed process %d (%s).\n", p->pid, p->comm);
148 * We give our sacrificial lamb high priority and access to
149 * all the memory it needs. That way it should be able to
150 * exit() and clear out its resources quickly...
153 p->flags |= PF_MEMALLOC | PF_MEMDIE;
155 /* This process has hardware access, be more careful. */
156 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO)) {
157 force_sig(SIGTERM, p);
159 force_sig(SIGKILL, p);
163 static struct mm_struct *oom_kill_task(task_t *p)
165 struct mm_struct *mm = get_task_mm(p);
166 if (!mm || mm == &init_mm)
174 * oom_kill - kill the "best" process when we run out of memory
176 * If we run out of memory, we have the choice between either
177 * killing a random task (bad), letting the system crash (worse)
178 * OR try to be smart about which process to kill. Note that we
179 * don't have to be perfect here, we just have to be good.
181 static void oom_kill(void)
183 struct mm_struct *mm;
184 struct task_struct *g, *p, *q;
186 read_lock(&tasklist_lock);
188 p = select_bad_process();
190 /* Found nothing?!?! Either we hang forever, or we panic. */
193 panic("Out of memory and no killable processes...\n");
196 mm = oom_kill_task(p);
200 * kill all processes that share the ->mm (i.e. all threads),
201 * but are in a different thread group
204 if (q->mm == mm && q->tgid != p->tgid)
206 while_each_thread(g, q);
208 printk(KERN_INFO "Fixed up OOM kill of mm-less task\n");
209 read_unlock(&tasklist_lock);
213 * Make kswapd go out of the way, so "p" has a good chance of
214 * killing itself before someone else gets the chance to ask
222 * out_of_memory - is the system out of memory?
224 void out_of_memory(int gfp_mask)
227 * oom_lock protects out_of_memory()'s static variables.
228 * It's a global lock; this is not performance-critical.
230 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
231 static unsigned long first, last, count, lastkill;
232 unsigned long now, since;
234 spin_lock(&oom_lock);
240 * If it's been a long time since last failure,
247 * If we haven't tried for at least one second,
248 * we're not really oom.
255 * If we have gotten only a few failures,
256 * we're not really oom.
262 * If we just killed a process, wait a while
263 * to give that task a chance to exit. This
264 * avoids killing multiple processes needlessly.
266 since = now - lastkill;
271 * Ok, really out of memory. Kill something.
275 printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
278 /* oom_kill() sleeps */
279 spin_unlock(&oom_lock);
281 spin_lock(&oom_lock);
285 * We dropped the lock above, so check to be sure the variable
286 * first only ever increases to prevent false OOM's.
288 if (time_after(now, first))
293 spin_unlock(&oom_lock);