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>
28 * oom_badness - calculate a numeric value for how bad this task has been
29 * @p: task struct of which task we should calculate
30 * @uptime: current uptime in seconds
32 * The formula used is relatively simple and documented inline in the
33 * function. The main rationale is that we want to select a good task
34 * to kill when we run out of memory.
36 * Good in this context means that:
37 * 1) we lose the minimum amount of work done
38 * 2) we recover a large amount of memory
39 * 3) we don't kill anything innocent of eating tons of memory
40 * 4) we want to kill the minimum amount of processes (one)
41 * 5) we try to kill the process the user expects us to kill, this
42 * algorithm has been meticulously tuned to meet the principle
43 * of least surprise ... (be careful when you change it)
46 unsigned long badness(struct task_struct *p, unsigned long uptime)
48 unsigned long points, cpu_time, run_time, s;
49 struct list_head *tsk;
55 * The memory size of the process is the basis for the badness.
57 points = p->mm->total_vm;
58 /* FIXME: add vserver badness ;) */
61 * Processes which fork a lot of child processes are likely
62 * a good choice. We add half the vmsize of the children if they
63 * have an own mm. This prevents forking servers to flood the
64 * machine with an endless amount of children. In case a single
65 * child is eating the vast majority of memory, adding only half
66 * to the parents will make the child our kill candidate of choice.
68 list_for_each(tsk, &p->children) {
69 struct task_struct *chld;
70 chld = list_entry(tsk, struct task_struct, sibling);
71 if (chld->mm != p->mm && chld->mm)
72 points += chld->mm->total_vm/2 + 1;
76 * CPU time is in tens of seconds and run time is in thousands
77 * of seconds. There is no particular reason for this other than
78 * that it turned out to work very well in practice.
80 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
83 if (uptime >= p->start_time.tv_sec)
84 run_time = (uptime - p->start_time.tv_sec) >> 10;
88 s = int_sqrt(cpu_time);
91 s = int_sqrt(int_sqrt(run_time));
96 * Niced processes are most likely less important, so double
97 * their badness points.
103 * Superuser processes are usually more important, so we make it
104 * less likely that we kill those.
106 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
107 p->uid == 0 || p->euid == 0)
111 * We don't want to kill a process with direct hardware access.
112 * Not only could that mess up the hardware, but usually users
113 * tend to only have this flag set on applications they think
116 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
120 * Adjust the score by oomkilladj.
123 if (p->oomkilladj > 0)
124 points <<= p->oomkilladj;
126 points >>= -(p->oomkilladj);
130 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
131 p->pid, p->comm, points);
137 * Types of limitations to the nodes from which allocations may occur
139 #define CONSTRAINT_NONE 1
140 #define CONSTRAINT_MEMORY_POLICY 2
141 #define CONSTRAINT_CPUSET 3
144 * Determine the type of allocation constraint.
146 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
150 nodemask_t nodes = node_online_map;
152 for (z = zonelist->zones; *z; z++)
153 if (cpuset_zone_allowed(*z, gfp_mask))
154 node_clear((*z)->zone_pgdat->node_id,
157 return CONSTRAINT_CPUSET;
159 if (!nodes_empty(nodes))
160 return CONSTRAINT_MEMORY_POLICY;
163 return CONSTRAINT_NONE;
167 * Simple selection loop. We chose the process with the highest
168 * number of 'points'. We expect the caller will lock the tasklist.
170 * (not docbooked, we don't want this one cluttering up the manual)
172 static struct task_struct *select_bad_process(unsigned long *ppoints)
174 struct task_struct *g, *p;
175 struct task_struct *chosen = NULL;
176 struct timespec uptime;
179 do_posix_clock_monotonic_gettime(&uptime);
180 do_each_thread(g, p) {
181 unsigned long points;
184 /* skip the init task with pid == 1 */
187 if (p->oomkilladj == OOM_DISABLE)
189 /* If p's nodes don't overlap ours, it won't help to kill p. */
190 if (!cpuset_excl_nodes_overlap(p))
194 * This is in the process of releasing memory so for wait it
195 * to finish before killing some other task by mistake.
197 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
198 p->flags & PF_EXITING;
199 if (releasing && !(p->flags & PF_DEAD))
200 return ERR_PTR(-1UL);
201 if (p->flags & PF_SWAPOFF)
204 points = badness(p, uptime.tv_sec);
205 if (points > *ppoints || !chosen) {
209 } while_each_thread(g, p);
214 * We must be careful though to never send SIGKILL a process with
215 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
216 * we select a process with CAP_SYS_RAW_IO set).
218 static void __oom_kill_task(task_t *p, const char *message)
222 printk(KERN_WARNING "tried to kill init!\n");
227 if (!p->mm || p->mm == &init_mm) {
229 printk(KERN_WARNING "tried to kill an mm-less task!\n");
234 printk(KERN_ERR "%s: Killed process %d (%s).\n",
235 message, p->pid, p->comm);
238 * We give our sacrificial lamb high priority and access to
239 * all the memory it needs. That way it should be able to
240 * exit() and clear out its resources quickly...
243 set_tsk_thread_flag(p, TIF_MEMDIE);
245 force_sig(SIGKILL, p);
248 static struct mm_struct *oom_kill_task(task_t *p, const char *message)
250 struct mm_struct *mm = get_task_mm(p);
255 if (mm == &init_mm) {
260 __oom_kill_task(p, message);
262 * kill all processes that share the ->mm (i.e. all threads),
263 * but are in a different thread group
266 if (q->mm == mm && q->tgid != p->tgid)
267 __oom_kill_task(q, message);
268 while_each_thread(g, q);
273 static struct mm_struct *oom_kill_process(struct task_struct *p,
274 unsigned long points, const char *message)
276 struct mm_struct *mm;
277 struct task_struct *c;
278 struct list_head *tsk;
280 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
281 "children.\n", p->pid, p->comm, points);
282 /* Try to kill a child first */
283 list_for_each(tsk, &p->children) {
284 c = list_entry(tsk, struct task_struct, sibling);
287 mm = oom_kill_task(c, message);
291 return oom_kill_task(p, message);
295 * oom_kill - kill the "best" process when we run out of memory
297 * If we run out of memory, we have the choice between either
298 * killing a random task (bad), letting the system crash (worse)
299 * OR try to be smart about which process to kill. Note that we
300 * don't have to be perfect here, we just have to be good.
302 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
304 struct mm_struct *mm = NULL;
306 unsigned long points = 0;
308 if (printk_ratelimit()) {
309 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
316 read_lock(&tasklist_lock);
319 * Check if there were limitations on the allocation (only relevant for
320 * NUMA) that may require different handling.
322 switch (constrained_alloc(zonelist, gfp_mask)) {
323 case CONSTRAINT_MEMORY_POLICY:
324 mm = oom_kill_process(current, points,
325 "No available memory (MPOL_BIND)");
328 case CONSTRAINT_CPUSET:
329 mm = oom_kill_process(current, points,
330 "No available memory in cpuset");
333 case CONSTRAINT_NONE:
336 * Rambo mode: Shoot down a process and hope it solves whatever
337 * issues we may have.
339 p = select_bad_process(&points);
341 if (PTR_ERR(p) == -1UL)
344 /* Found nothing?!?! Either we hang forever, or we panic. */
346 read_unlock(&tasklist_lock);
348 panic("Out of memory and no killable processes...\n");
351 mm = oom_kill_process(p, points, "Out of memory");
359 read_unlock(&tasklist_lock);
365 * Give "p" a good chance of killing itself before we
366 * retry to allocate memory unless "p" is current
368 if (!test_thread_flag(TIF_MEMDIE))
369 schedule_timeout_uninterruptible(1);