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;
50 struct task_struct *child;
60 * The memory size of the process is the basis for the badness.
62 points = mm->total_vm;
65 * After this unlock we can no longer dereference local variable `mm'
69 /* FIXME: add vserver badness ;) */
72 * Processes which fork a lot of child processes are likely
73 * a good choice. We add half the vmsize of the children if they
74 * have an own mm. This prevents forking servers to flood the
75 * machine with an endless amount of children. In case a single
76 * child is eating the vast majority of memory, adding only half
77 * to the parents will make the child our kill candidate of choice.
79 list_for_each_entry(child, &p->children, sibling) {
81 if (child->mm != mm && child->mm)
82 points += child->mm->total_vm/2 + 1;
87 * CPU time is in tens of seconds and run time is in thousands
88 * of seconds. There is no particular reason for this other than
89 * that it turned out to work very well in practice.
91 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
94 if (uptime >= p->start_time.tv_sec)
95 run_time = (uptime - p->start_time.tv_sec) >> 10;
99 s = int_sqrt(cpu_time);
102 s = int_sqrt(int_sqrt(run_time));
107 * Niced processes are most likely less important, so double
108 * their badness points.
110 if (task_nice(p) > 0)
114 * Superuser processes are usually more important, so we make it
115 * less likely that we kill those.
117 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
118 p->uid == 0 || p->euid == 0)
122 * We don't want to kill a process with direct hardware access.
123 * Not only could that mess up the hardware, but usually users
124 * tend to only have this flag set on applications they think
127 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
131 * Adjust the score by oomkilladj.
134 if (p->oomkilladj > 0)
135 points <<= p->oomkilladj;
137 points >>= -(p->oomkilladj);
141 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
142 p->pid, p->comm, points);
148 * Types of limitations to the nodes from which allocations may occur
150 #define CONSTRAINT_NONE 1
151 #define CONSTRAINT_MEMORY_POLICY 2
152 #define CONSTRAINT_CPUSET 3
155 * Determine the type of allocation constraint.
157 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
161 nodemask_t nodes = node_online_map;
163 for (z = zonelist->zones; *z; z++)
164 if (cpuset_zone_allowed(*z, gfp_mask))
165 node_clear((*z)->zone_pgdat->node_id,
168 return CONSTRAINT_CPUSET;
170 if (!nodes_empty(nodes))
171 return CONSTRAINT_MEMORY_POLICY;
174 return CONSTRAINT_NONE;
178 * Simple selection loop. We chose the process with the highest
179 * number of 'points'. We expect the caller will lock the tasklist.
181 * (not docbooked, we don't want this one cluttering up the manual)
183 static struct task_struct *select_bad_process(unsigned long *ppoints)
185 struct task_struct *g, *p;
186 struct task_struct *chosen = NULL;
187 struct timespec uptime;
190 do_posix_clock_monotonic_gettime(&uptime);
191 do_each_thread(g, p) {
192 unsigned long points;
195 /* skip the init task with pid == 1 */
198 if (p->oomkilladj == OOM_DISABLE)
200 /* If p's nodes don't overlap ours, it won't help to kill p. */
201 if (!cpuset_excl_nodes_overlap(p))
205 * This is in the process of releasing memory so for wait it
206 * to finish before killing some other task by mistake.
208 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
209 p->flags & PF_EXITING;
210 if (releasing && !(p->flags & PF_DEAD))
211 return ERR_PTR(-1UL);
212 if (p->flags & PF_SWAPOFF)
215 points = badness(p, uptime.tv_sec);
216 if (points > *ppoints || !chosen) {
220 } while_each_thread(g, p);
225 * We must be careful though to never send SIGKILL a process with
226 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
227 * we select a process with CAP_SYS_RAW_IO set).
229 static void __oom_kill_task(task_t *p, const char *message)
233 printk(KERN_WARNING "tried to kill init!\n");
238 if (!p->mm || p->mm == &init_mm) {
240 printk(KERN_WARNING "tried to kill an mm-less task!\n");
245 printk(KERN_ERR "%s: Killed process %d (%s).\n",
246 message, p->pid, p->comm);
249 * We give our sacrificial lamb high priority and access to
250 * all the memory it needs. That way it should be able to
251 * exit() and clear out its resources quickly...
254 set_tsk_thread_flag(p, TIF_MEMDIE);
256 force_sig(SIGKILL, p);
259 static int oom_kill_task(task_t *p, const char *message)
261 struct mm_struct *mm;
266 /* WARNING: mm may not be dereferenced since we did not obtain its
267 * value from get_task_mm(p). This is OK since all we need to do is
268 * compare mm to q->mm below.
270 * Furthermore, even if mm contains a non-NULL value, p->mm may
271 * change to NULL at any time since we do not hold task_lock(p).
272 * However, this is of no concern to us.
275 if (mm == NULL || mm == &init_mm)
278 __oom_kill_task(p, message);
280 * kill all processes that share the ->mm (i.e. all threads),
281 * but are in a different thread group
284 if (q->mm == mm && q->tgid != p->tgid)
285 __oom_kill_task(q, message);
286 while_each_thread(g, q);
291 static int oom_kill_process(struct task_struct *p, unsigned long points,
294 struct task_struct *c;
295 struct list_head *tsk;
297 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
298 "children.\n", p->pid, p->comm, points);
299 /* Try to kill a child first */
300 list_for_each(tsk, &p->children) {
301 c = list_entry(tsk, struct task_struct, sibling);
304 if (!oom_kill_task(c, message))
307 return oom_kill_task(p, message);
311 * oom_kill - kill the "best" process when we run out of memory
313 * If we run out of memory, we have the choice between either
314 * killing a random task (bad), letting the system crash (worse)
315 * OR try to be smart about which process to kill. Note that we
316 * don't have to be perfect here, we just have to be good.
318 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
321 unsigned long points = 0;
323 if (printk_ratelimit()) {
324 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
331 read_lock(&tasklist_lock);
334 * Check if there were limitations on the allocation (only relevant for
335 * NUMA) that may require different handling.
337 switch (constrained_alloc(zonelist, gfp_mask)) {
338 case CONSTRAINT_MEMORY_POLICY:
339 oom_kill_process(current, points,
340 "No available memory (MPOL_BIND)");
343 case CONSTRAINT_CPUSET:
344 oom_kill_process(current, points,
345 "No available memory in cpuset");
348 case CONSTRAINT_NONE:
351 * Rambo mode: Shoot down a process and hope it solves whatever
352 * issues we may have.
354 p = select_bad_process(&points);
356 if (PTR_ERR(p) == -1UL)
359 /* Found nothing?!?! Either we hang forever, or we panic. */
361 read_unlock(&tasklist_lock);
363 panic("Out of memory and no killable processes...\n");
366 if (oom_kill_process(p, points, "Out of memory"))
373 read_unlock(&tasklist_lock);
377 * Give "p" a good chance of killing itself before we
378 * retry to allocate memory unless "p" is current
380 if (!test_thread_flag(TIF_MEMDIE))
381 schedule_timeout_uninterruptible(1);