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/config.h>
20 #include <linux/sched.h>
21 #include <linux/swap.h>
22 #include <linux/timex.h>
23 #include <linux/jiffies.h>
24 #include <linux/cpuset.h>
29 * oom_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;
50 struct list_head *tsk;
56 * The memory size of the process is the basis for the badness.
58 points = p->mm->total_vm;
59 /* FIXME: add vserver badness ;) */
62 * Processes which fork a lot of child processes are likely
63 * a good choice. We add half the vmsize of the children if they
64 * have an own mm. This prevents forking servers to flood the
65 * machine with an endless amount of children. In case a single
66 * child is eating the vast majority of memory, adding only half
67 * to the parents will make the child our kill candidate of choice.
69 list_for_each(tsk, &p->children) {
70 struct task_struct *chld;
71 chld = list_entry(tsk, struct task_struct, sibling);
72 if (chld->mm != p->mm && chld->mm)
73 points += chld->mm->total_vm/2 + 1;
77 * CPU time is in tens of seconds and run time is in thousands
78 * of seconds. There is no particular reason for this other than
79 * that it turned out to work very well in practice.
81 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
84 if (uptime >= p->start_time.tv_sec)
85 run_time = (uptime - p->start_time.tv_sec) >> 10;
89 s = int_sqrt(cpu_time);
92 s = int_sqrt(int_sqrt(run_time));
97 * Niced processes are most likely less important, so double
98 * their badness points.
100 if (task_nice(p) > 0)
104 * Superuser processes are usually more important, so we make it
105 * less likely that we kill those.
107 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
108 p->uid == 0 || p->euid == 0)
112 * We don't want to kill a process with direct hardware access.
113 * Not only could that mess up the hardware, but usually users
114 * tend to only have this flag set on applications they think
117 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
121 * Adjust the score by oomkilladj.
124 if (p->oomkilladj > 0)
125 points <<= p->oomkilladj;
127 points >>= -(p->oomkilladj);
131 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
132 p->pid, p->comm, points);
137 #if defined(CONFIG_OOM_PANIC) && defined(CONFIG_OOM_KILLER)
138 #warning Only define OOM_PANIC or OOM_KILLER; not both
141 #ifdef CONFIG_OOM_KILLER
143 * Types of limitations to the nodes from which allocations may occur
145 #define CONSTRAINT_NONE 1
146 #define CONSTRAINT_MEMORY_POLICY 2
147 #define CONSTRAINT_CPUSET 3
150 * Determine the type of allocation constraint.
152 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
156 nodemask_t nodes = node_online_map;
158 for (z = zonelist->zones; *z; z++)
159 if (cpuset_zone_allowed(*z, gfp_mask))
160 node_clear((*z)->zone_pgdat->node_id,
163 return CONSTRAINT_CPUSET;
165 if (!nodes_empty(nodes))
166 return CONSTRAINT_MEMORY_POLICY;
169 return CONSTRAINT_NONE;
173 * Simple selection loop. We chose the process with the highest
174 * number of 'points'. We expect the caller will lock the tasklist.
176 * (not docbooked, we don't want this one cluttering up the manual)
178 static struct task_struct *select_bad_process(unsigned long *ppoints)
180 struct task_struct *g, *p;
181 struct task_struct *chosen = NULL;
182 struct timespec uptime;
185 do_posix_clock_monotonic_gettime(&uptime);
186 do_each_thread(g, p) {
187 unsigned long points;
190 /* skip the init task with pid == 1 */
193 if (p->oomkilladj == OOM_DISABLE)
195 /* If p's nodes don't overlap ours, it won't help to kill p. */
196 if (!cpuset_excl_nodes_overlap(p))
200 * This is in the process of releasing memory so for wait it
201 * to finish before killing some other task by mistake.
203 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
204 p->flags & PF_EXITING;
205 if (releasing && !(p->flags & PF_DEAD))
206 return ERR_PTR(-1UL);
207 if (p->flags & PF_SWAPOFF)
210 points = badness(p, uptime.tv_sec);
211 if (points > *ppoints || !chosen) {
215 } while_each_thread(g, p);
220 * We must be careful though to never send SIGKILL a process with
221 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
222 * we select a process with CAP_SYS_RAW_IO set).
224 static void __oom_kill_task(task_t *p, const char *message)
228 printk(KERN_WARNING "tried to kill init!\n");
233 if (!p->mm || p->mm == &init_mm) {
235 printk(KERN_WARNING "tried to kill an mm-less task!\n");
240 printk(KERN_ERR "%s: Killed process %d (%s).\n",
241 message, p->pid, p->comm);
244 * We give our sacrificial lamb high priority and access to
245 * all the memory it needs. That way it should be able to
246 * exit() and clear out its resources quickly...
249 set_tsk_thread_flag(p, TIF_MEMDIE);
251 force_sig(SIGKILL, p);
254 static struct mm_struct *oom_kill_task(task_t *p, const char *message)
256 struct mm_struct *mm = get_task_mm(p);
261 if (mm == &init_mm) {
266 __oom_kill_task(p, message);
268 * kill all processes that share the ->mm (i.e. all threads),
269 * but are in a different thread group
272 if (q->mm == mm && q->tgid != p->tgid)
273 __oom_kill_task(q, message);
274 while_each_thread(g, q);
279 static struct mm_struct *oom_kill_process(struct task_struct *p,
280 unsigned long points, const char *message)
282 struct mm_struct *mm;
283 struct task_struct *c;
284 struct list_head *tsk;
286 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
287 "children.\n", p->pid, p->comm, points);
288 /* Try to kill a child first */
289 list_for_each(tsk, &p->children) {
290 c = list_entry(tsk, struct task_struct, sibling);
293 mm = oom_kill_task(c, message);
297 return oom_kill_task(p, message);
301 * oom_kill - kill the "best" process when we run out of memory
303 * If we run out of memory, we have the choice between either
304 * killing a random task (bad), letting the system crash (worse)
305 * OR try to be smart about which process to kill. Note that we
306 * don't have to be perfect here, we just have to be good.
308 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
310 struct mm_struct *mm = NULL;
312 unsigned long points = 0;
314 if (printk_ratelimit()) {
315 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
322 read_lock(&tasklist_lock);
325 * Check if there were limitations on the allocation (only relevant for
326 * NUMA) that may require different handling.
328 switch (constrained_alloc(zonelist, gfp_mask)) {
329 case CONSTRAINT_MEMORY_POLICY:
330 mm = oom_kill_process(current, points,
331 "No available memory (MPOL_BIND)");
334 case CONSTRAINT_CPUSET:
335 mm = oom_kill_process(current, points,
336 "No available memory in cpuset");
339 case CONSTRAINT_NONE:
342 * Rambo mode: Shoot down a process and hope it solves whatever
343 * issues we may have.
345 p = select_bad_process(&points);
347 if (PTR_ERR(p) == -1UL)
350 /* Found nothing?!?! Either we hang forever, or we panic. */
352 read_unlock(&tasklist_lock);
354 panic("Out of memory and no killable processes...\n");
357 mm = oom_kill_process(p, points, "Out of memory");
365 read_unlock(&tasklist_lock);
371 * Give "p" a good chance of killing itself before we
372 * retry to allocate memory unless "p" is current
374 if (!test_thread_flag(TIF_MEMDIE))
375 schedule_timeout_uninterruptible(1);
377 #endif /* CONFIG_OOM_KILLER */
379 #ifdef CONFIG_OOM_PANIC
381 * out_of_memory - panic if the system out of memory?
383 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
386 * oom_lock protects out_of_memory()'s static variables.
387 * It's a global lock; this is not performance-critical.
389 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
390 static unsigned long count;
392 spin_lock(&oom_lock);
395 * If we have gotten only a few failures,
396 * we're not really oom.
400 * Ok, really out of memory. Panic.
403 printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
406 panic("Out Of Memory");
408 spin_unlock(&oom_lock);
410 #endif /* CONFIG_OOM_PANIC */