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
29 * @p: current uptime in seconds
31 * The formula used is relatively simple and documented inline in the
32 * function. The main rationale is that we want to select a good task
33 * to kill when we run out of memory.
35 * Good in this context means that:
36 * 1) we lose the minimum amount of work done
37 * 2) we recover a large amount of memory
38 * 3) we don't kill anything innocent of eating tons of memory
39 * 4) we want to kill the minimum amount of processes (one)
40 * 5) we try to kill the process the user expects us to kill, this
41 * algorithm has been meticulously tuned to meet the principle
42 * of least surprise ... (be careful when you change it)
45 static unsigned long badness(struct task_struct *p, unsigned long uptime)
47 unsigned long points, cpu_time, run_time, s;
52 if (p->flags & PF_MEMDIE)
55 * The memory size of the process is the basis for the badness.
57 points = p->mm->total_vm;
58 /* add vserver badness ;) */
61 * CPU time is in tens of seconds and run time is in thousands
62 * of seconds. There is no particular reason for this other than
63 * that it turned out to work very well in practice.
65 cpu_time = (p->utime + p->stime) >> (SHIFT_HZ + 3);
67 if (uptime >= p->start_time.tv_sec)
68 run_time = (uptime - p->start_time.tv_sec) >> 10;
72 s = int_sqrt(cpu_time);
75 s = int_sqrt(int_sqrt(run_time));
80 * Niced processes are most likely less important, so double
81 * their badness points.
87 * Superuser processes are usually more important, so we make it
88 * less likely that we kill those.
90 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
91 p->uid == 0 || p->euid == 0)
95 * We don't want to kill a process with direct hardware access.
96 * Not only could that mess up the hardware, but usually users
97 * tend to only have this flag set on applications they think
100 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
103 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
104 p->pid, p->comm, points);
110 * Simple selection loop. We chose the process with the highest
111 * number of 'points'. We expect the caller will lock the tasklist.
113 * (not docbooked, we don't want this one cluttering up the manual)
115 static struct task_struct * select_bad_process(void)
117 unsigned long maxpoints = 0;
118 struct task_struct *g, *p;
119 struct task_struct *chosen = NULL;
120 struct timespec uptime;
122 do_posix_clock_monotonic_gettime(&uptime);
125 unsigned long points = badness(p, uptime.tv_sec);
126 if (points > maxpoints) {
130 if (p->flags & PF_SWAPOFF)
133 while_each_thread(g, p);
138 * We must be careful though to never send SIGKILL a process with
139 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
140 * we select a process with CAP_SYS_RAW_IO set).
142 static void __oom_kill_task(task_t *p)
145 if (!p->mm || p->mm == &init_mm) {
147 printk(KERN_WARNING "tried to kill an mm-less task!\n");
152 printk(KERN_ERR "Out of Memory: Killed process %d (%s).\n", p->pid, p->comm);
155 * We give our sacrificial lamb high priority and access to
156 * all the memory it needs. That way it should be able to
157 * exit() and clear out its resources quickly...
160 p->flags |= PF_MEMALLOC | PF_MEMDIE;
162 /* This process has hardware access, be more careful. */
163 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO)) {
164 force_sig(SIGTERM, p);
166 force_sig(SIGKILL, p);
170 static struct mm_struct *oom_kill_task(task_t *p)
172 struct mm_struct *mm = get_task_mm(p);
173 if (!mm || mm == &init_mm)
181 * oom_kill - kill the "best" process when we run out of memory
183 * If we run out of memory, we have the choice between either
184 * killing a random task (bad), letting the system crash (worse)
185 * OR try to be smart about which process to kill. Note that we
186 * don't have to be perfect here, we just have to be good.
188 static void oom_kill(void)
190 struct mm_struct *mm;
191 struct task_struct *g, *p, *q;
193 /* print the memory stats whenever we OOM kill */
196 read_lock(&tasklist_lock);
198 p = select_bad_process();
200 /* Found nothing?!?! Either we hang forever, or we panic. */
203 panic("Out of memory and no killable processes...\n");
206 mm = oom_kill_task(p);
210 * kill all processes that share the ->mm (i.e. all threads),
211 * but are in a different thread group
214 if (q->mm == mm && q->tgid != p->tgid)
216 while_each_thread(g, q);
218 printk(KERN_INFO "Fixed up OOM kill of mm-less task\n");
219 read_unlock(&tasklist_lock);
223 * Make kswapd go out of the way, so "p" has a good chance of
224 * killing itself before someone else gets the chance to ask
232 * out_of_memory - is the system out of memory?
234 void out_of_memory(int gfp_mask)
237 * oom_lock protects out_of_memory()'s static variables.
238 * It's a global lock; this is not performance-critical.
240 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
241 static unsigned long first, last, count, lastkill;
242 unsigned long now, since;
244 spin_lock(&oom_lock);
250 * If it's been a long time since last failure,
257 * If we haven't tried for at least one second,
258 * we're not really oom.
265 * If we have gotten only a few failures,
266 * we're not really oom.
272 * If we just killed a process, wait a while
273 * to give that task a chance to exit. This
274 * avoids killing multiple processes needlessly.
276 since = now - lastkill;
281 * Ok, really out of memory. Kill something.
285 printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
287 /* oom_kill() sleeps */
288 spin_unlock(&oom_lock);
290 spin_lock(&oom_lock);
294 * We dropped the lock above, so check to be sure the variable
295 * first only ever increases to prevent false OOM's.
297 if (time_after(now, first))
302 spin_unlock(&oom_lock);