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;
51 struct task_struct *child;
61 * The memory size of the process is the basis for the badness.
63 points = mm->total_vm;
66 * After this unlock we can no longer dereference local variable `mm'
70 /* FIXME: add vserver badness ;) */
73 * Processes which fork a lot of child processes are likely
74 * a good choice. We add half the vmsize of the children if they
75 * have an own mm. This prevents forking servers to flood the
76 * machine with an endless amount of children. In case a single
77 * child is eating the vast majority of memory, adding only half
78 * to the parents will make the child our kill candidate of choice.
80 list_for_each_entry(child, &p->children, sibling) {
82 if (child->mm != mm && child->mm)
83 points += child->mm->total_vm/2 + 1;
88 * CPU time is in tens of seconds and run time is in thousands
89 * of seconds. There is no particular reason for this other than
90 * that it turned out to work very well in practice.
92 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
95 if (uptime >= p->start_time.tv_sec)
96 run_time = (uptime - p->start_time.tv_sec) >> 10;
100 s = int_sqrt(cpu_time);
103 s = int_sqrt(int_sqrt(run_time));
108 * Niced processes are most likely less important, so double
109 * their badness points.
111 if (task_nice(p) > 0)
115 * Superuser processes are usually more important, so we make it
116 * less likely that we kill those.
118 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
119 p->uid == 0 || p->euid == 0)
123 * We don't want to kill a process with direct hardware access.
124 * Not only could that mess up the hardware, but usually users
125 * tend to only have this flag set on applications they think
128 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
132 * Adjust the score by oomkilladj.
135 if (p->oomkilladj > 0)
136 points <<= p->oomkilladj;
138 points >>= -(p->oomkilladj);
142 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
143 p->pid, p->comm, points);
148 #if defined(CONFIG_OOM_PANIC) && defined(CONFIG_OOM_KILLER)
149 #warning Only define OOM_PANIC or OOM_KILLER; not both
152 #ifdef CONFIG_OOM_KILLER
154 * Types of limitations to the nodes from which allocations may occur
156 #define CONSTRAINT_NONE 1
157 #define CONSTRAINT_MEMORY_POLICY 2
158 #define CONSTRAINT_CPUSET 3
161 * Determine the type of allocation constraint.
163 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
167 nodemask_t nodes = node_online_map;
169 for (z = zonelist->zones; *z; z++)
170 if (cpuset_zone_allowed(*z, gfp_mask))
171 node_clear((*z)->zone_pgdat->node_id,
174 return CONSTRAINT_CPUSET;
176 if (!nodes_empty(nodes))
177 return CONSTRAINT_MEMORY_POLICY;
180 return CONSTRAINT_NONE;
184 * Simple selection loop. We chose the process with the highest
185 * number of 'points'. We expect the caller will lock the tasklist.
187 * (not docbooked, we don't want this one cluttering up the manual)
189 static struct task_struct *select_bad_process(unsigned long *ppoints)
191 struct task_struct *g, *p;
192 struct task_struct *chosen = NULL;
193 struct timespec uptime;
196 do_posix_clock_monotonic_gettime(&uptime);
197 do_each_thread(g, p) {
198 unsigned long points;
201 /* skip the init task with pid == 1 */
204 if (p->oomkilladj == OOM_DISABLE)
206 /* If p's nodes don't overlap ours, it won't help to kill p. */
207 if (!cpuset_excl_nodes_overlap(p))
211 * This is in the process of releasing memory so for wait it
212 * to finish before killing some other task by mistake.
214 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
215 p->flags & PF_EXITING;
216 if (releasing && !(p->flags & PF_DEAD))
217 return ERR_PTR(-1UL);
218 if (p->flags & PF_SWAPOFF)
221 points = badness(p, uptime.tv_sec);
222 if (points > *ppoints || !chosen) {
226 } while_each_thread(g, p);
231 * We must be careful though to never send SIGKILL a process with
232 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
233 * we select a process with CAP_SYS_RAW_IO set).
235 static void __oom_kill_task(task_t *p, const char *message)
239 printk(KERN_WARNING "tried to kill init!\n");
244 if (!p->mm || p->mm == &init_mm) {
246 printk(KERN_WARNING "tried to kill an mm-less task!\n");
251 printk(KERN_ERR "%s: Killed process %d (%s).\n",
252 message, p->pid, p->comm);
255 * We give our sacrificial lamb high priority and access to
256 * all the memory it needs. That way it should be able to
257 * exit() and clear out its resources quickly...
260 set_tsk_thread_flag(p, TIF_MEMDIE);
262 force_sig(SIGKILL, p);
265 static int oom_kill_task(task_t *p, const char *message)
267 struct mm_struct *mm;
272 /* WARNING: mm may not be dereferenced since we did not obtain its
273 * value from get_task_mm(p). This is OK since all we need to do is
274 * compare mm to q->mm below.
276 * Furthermore, even if mm contains a non-NULL value, p->mm may
277 * change to NULL at any time since we do not hold task_lock(p).
278 * However, this is of no concern to us.
281 if (mm == NULL || mm == &init_mm)
284 __oom_kill_task(p, message);
286 * kill all processes that share the ->mm (i.e. all threads),
287 * but are in a different thread group
290 if (q->mm == mm && q->tgid != p->tgid)
291 __oom_kill_task(q, message);
292 while_each_thread(g, q);
297 static int oom_kill_process(struct task_struct *p, unsigned long points,
300 struct task_struct *c;
301 struct list_head *tsk;
303 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
304 "children.\n", p->pid, p->comm, points);
305 /* Try to kill a child first */
306 list_for_each(tsk, &p->children) {
307 c = list_entry(tsk, struct task_struct, sibling);
310 if (!oom_kill_task(c, message))
313 return oom_kill_task(p, message);
317 * oom_kill - kill the "best" process when we run out of memory
319 * If we run out of memory, we have the choice between either
320 * killing a random task (bad), letting the system crash (worse)
321 * OR try to be smart about which process to kill. Note that we
322 * don't have to be perfect here, we just have to be good.
324 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
327 unsigned long points = 0;
329 if (printk_ratelimit()) {
330 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
337 read_lock(&tasklist_lock);
340 * Check if there were limitations on the allocation (only relevant for
341 * NUMA) that may require different handling.
343 switch (constrained_alloc(zonelist, gfp_mask)) {
344 case CONSTRAINT_MEMORY_POLICY:
345 oom_kill_process(current, points,
346 "No available memory (MPOL_BIND)");
349 case CONSTRAINT_CPUSET:
350 oom_kill_process(current, points,
351 "No available memory in cpuset");
354 case CONSTRAINT_NONE:
357 * Rambo mode: Shoot down a process and hope it solves whatever
358 * issues we may have.
360 p = select_bad_process(&points);
362 if (PTR_ERR(p) == -1UL)
365 /* Found nothing?!?! Either we hang forever, or we panic. */
367 read_unlock(&tasklist_lock);
369 panic("Out of memory and no killable processes...\n");
372 if (oom_kill_process(p, points, "Out of memory"))
379 read_unlock(&tasklist_lock);
383 * Give "p" a good chance of killing itself before we
384 * retry to allocate memory unless "p" is current
386 if (!test_thread_flag(TIF_MEMDIE))
387 schedule_timeout_uninterruptible(1);
389 #endif /* CONFIG_OOM_KILLER */
391 #ifdef CONFIG_OOM_PANIC
393 * out_of_memory - panic if the system out of memory?
395 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
398 * oom_lock protects out_of_memory()'s static variables.
399 * It's a global lock; this is not performance-critical.
401 static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
402 static unsigned long count;
404 spin_lock(&oom_lock);
407 * If we have gotten only a few failures,
408 * we're not really oom.
412 * Ok, really out of memory. Panic.
415 printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
418 panic("Out Of Memory");
420 spin_unlock(&oom_lock);
422 #endif /* CONFIG_OOM_PANIC */