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/oom.h>
20 #include <linux/err.h>
21 #include <linux/sched.h>
22 #include <linux/swap.h>
23 #include <linux/timex.h>
24 #include <linux/jiffies.h>
25 #include <linux/cpuset.h>
26 #include <linux/module.h>
27 #include <linux/notifier.h>
29 int sysctl_panic_on_oom;
30 static DEFINE_MUTEX(zone_scan_mutex);
34 * badness - calculate a numeric value for how bad this task has been
35 * @p: task struct of which task we should calculate
36 * @uptime: current uptime in seconds
38 * The formula used is relatively simple and documented inline in the
39 * function. The main rationale is that we want to select a good task
40 * to kill when we run out of memory.
42 * Good in this context means that:
43 * 1) we lose the minimum amount of work done
44 * 2) we recover a large amount of memory
45 * 3) we don't kill anything innocent of eating tons of memory
46 * 4) we want to kill the minimum amount of processes (one)
47 * 5) we try to kill the process the user expects us to kill, this
48 * algorithm has been meticulously tuned to meet the principle
49 * of least surprise ... (be careful when you change it)
52 unsigned long badness(struct task_struct *p, unsigned long uptime)
54 unsigned long points, cpu_time, run_time, s;
56 struct task_struct *child;
66 * The memory size of the process is the basis for the badness.
68 points = mm->total_vm;
71 * After this unlock we can no longer dereference local variable `mm'
76 * swapoff can easily use up all memory, so kill those first.
78 if (p->flags & PF_SWAPOFF)
82 * Processes which fork a lot of child processes are likely
83 * a good choice. We add half the vmsize of the children if they
84 * have an own mm. This prevents forking servers to flood the
85 * machine with an endless amount of children. In case a single
86 * child is eating the vast majority of memory, adding only half
87 * to the parents will make the child our kill candidate of choice.
89 list_for_each_entry(child, &p->children, sibling) {
91 if (child->mm != mm && child->mm)
92 points += child->mm->total_vm/2 + 1;
97 * CPU time is in tens of seconds and run time is in thousands
98 * of seconds. There is no particular reason for this other than
99 * that it turned out to work very well in practice.
101 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
104 if (uptime >= p->start_time.tv_sec)
105 run_time = (uptime - p->start_time.tv_sec) >> 10;
109 s = int_sqrt(cpu_time);
112 s = int_sqrt(int_sqrt(run_time));
117 * Niced processes are most likely less important, so double
118 * their badness points.
120 if (task_nice(p) > 0)
124 * Superuser processes are usually more important, so we make it
125 * less likely that we kill those.
127 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
128 p->uid == 0 || p->euid == 0)
132 * We don't want to kill a process with direct hardware access.
133 * Not only could that mess up the hardware, but usually users
134 * tend to only have this flag set on applications they think
137 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
141 * If p's nodes don't overlap ours, it may still help to kill p
142 * because p may have allocated or otherwise mapped memory on
143 * this node before. However it will be less likely.
145 if (!cpuset_excl_nodes_overlap(p))
149 * Adjust the score by oomkilladj.
152 if (p->oomkilladj > 0) {
155 points <<= p->oomkilladj;
157 points >>= -(p->oomkilladj);
161 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
162 p->pid, p->comm, points);
168 * Determine the type of allocation constraint.
170 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
175 nodemask_t nodes = node_states[N_HIGH_MEMORY];
177 for (z = zonelist->zones; *z; z++)
178 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
179 node_clear(zone_to_nid(*z), nodes);
181 return CONSTRAINT_CPUSET;
183 if (!nodes_empty(nodes))
184 return CONSTRAINT_MEMORY_POLICY;
187 return CONSTRAINT_NONE;
191 * Simple selection loop. We chose the process with the highest
192 * number of 'points'. We expect the caller will lock the tasklist.
194 * (not docbooked, we don't want this one cluttering up the manual)
196 static struct task_struct *select_bad_process(unsigned long *ppoints)
198 struct task_struct *g, *p;
199 struct task_struct *chosen = NULL;
200 struct timespec uptime;
203 do_posix_clock_monotonic_gettime(&uptime);
204 do_each_thread(g, p) {
205 unsigned long points;
208 * skip kernel threads and tasks which have already released
213 /* skip the init task */
218 * This task already has access to memory reserves and is
219 * being killed. Don't allow any other task access to the
222 * Note: this may have a chance of deadlock if it gets
223 * blocked waiting for another task which itself is waiting
224 * for memory. Is there a better alternative?
226 if (test_tsk_thread_flag(p, TIF_MEMDIE))
227 return ERR_PTR(-1UL);
230 * This is in the process of releasing memory so wait for it
231 * to finish before killing some other task by mistake.
233 * However, if p is the current task, we allow the 'kill' to
234 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
235 * which will allow it to gain access to memory reserves in
236 * the process of exiting and releasing its resources.
237 * Otherwise we could get an easy OOM deadlock.
239 if (p->flags & PF_EXITING) {
241 return ERR_PTR(-1UL);
244 *ppoints = ULONG_MAX;
247 if (p->oomkilladj == OOM_DISABLE)
250 points = badness(p, uptime.tv_sec);
251 if (points > *ppoints || !chosen) {
255 } while_each_thread(g, p);
261 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
262 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
265 static void __oom_kill_task(struct task_struct *p, int verbose)
269 printk(KERN_WARNING "tried to kill init!\n");
275 printk(KERN_WARNING "tried to kill an mm-less task!\n");
280 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm);
283 * We give our sacrificial lamb high priority and access to
284 * all the memory it needs. That way it should be able to
285 * exit() and clear out its resources quickly...
288 set_tsk_thread_flag(p, TIF_MEMDIE);
290 force_sig(SIGKILL, p);
293 static int oom_kill_task(struct task_struct *p)
295 struct mm_struct *mm;
296 struct task_struct *g, *q;
300 /* WARNING: mm may not be dereferenced since we did not obtain its
301 * value from get_task_mm(p). This is OK since all we need to do is
302 * compare mm to q->mm below.
304 * Furthermore, even if mm contains a non-NULL value, p->mm may
305 * change to NULL at any time since we do not hold task_lock(p).
306 * However, this is of no concern to us.
313 * Don't kill the process if any threads are set to OOM_DISABLE
315 do_each_thread(g, q) {
316 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
318 } while_each_thread(g, q);
320 __oom_kill_task(p, 1);
323 * kill all processes that share the ->mm (i.e. all threads),
324 * but are in a different thread group. Don't let them have access
325 * to memory reserves though, otherwise we might deplete all memory.
327 do_each_thread(g, q) {
328 if (q->mm == mm && q->tgid != p->tgid)
329 force_sig(SIGKILL, q);
330 } while_each_thread(g, q);
335 static int oom_kill_process(struct task_struct *p, unsigned long points,
338 struct task_struct *c;
339 struct list_head *tsk;
342 * If the task is already exiting, don't alarm the sysadmin or kill
343 * its children or threads, just set TIF_MEMDIE so it can die quickly
345 if (p->flags & PF_EXITING) {
346 __oom_kill_task(p, 0);
350 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
351 message, p->pid, p->comm, points);
353 /* Try to kill a child first */
354 list_for_each(tsk, &p->children) {
355 c = list_entry(tsk, struct task_struct, sibling);
358 if (!oom_kill_task(c))
361 return oom_kill_task(p);
364 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
366 int register_oom_notifier(struct notifier_block *nb)
368 return blocking_notifier_chain_register(&oom_notify_list, nb);
370 EXPORT_SYMBOL_GPL(register_oom_notifier);
372 int unregister_oom_notifier(struct notifier_block *nb)
374 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
376 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
379 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
380 * if a parallel OOM killing is already taking place that includes a zone in
381 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
383 int try_set_zone_oom(struct zonelist *zonelist)
390 mutex_lock(&zone_scan_mutex);
392 if (zone_is_oom_locked(*z)) {
396 } while (*(++z) != NULL);
399 * Lock each zone in the zonelist under zone_scan_mutex so a parallel
400 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
404 zone_set_flag(*z, ZONE_OOM_LOCKED);
405 } while (*(++z) != NULL);
407 mutex_unlock(&zone_scan_mutex);
412 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
413 * allocation attempts with zonelists containing them may now recall the OOM
414 * killer, if necessary.
416 void clear_zonelist_oom(struct zonelist *zonelist)
422 mutex_lock(&zone_scan_mutex);
424 zone_clear_flag(*z, ZONE_OOM_LOCKED);
425 } while (*(++z) != NULL);
426 mutex_unlock(&zone_scan_mutex);
430 * out_of_memory - kill the "best" process when we run out of memory
432 * If we run out of memory, we have the choice between either
433 * killing a random task (bad), letting the system crash (worse)
434 * OR try to be smart about which process to kill. Note that we
435 * don't have to be perfect here, we just have to be good.
437 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
439 struct task_struct *p;
440 unsigned long points = 0;
441 unsigned long freed = 0;
442 enum oom_constraint constraint;
444 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
446 /* Got some memory back in the last second. */
449 if (printk_ratelimit()) {
450 printk(KERN_WARNING "%s invoked oom-killer: "
451 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
452 current->comm, gfp_mask, order, current->oomkilladj);
457 if (sysctl_panic_on_oom == 2)
458 panic("out of memory. Compulsory panic_on_oom is selected.\n");
461 * Check if there were limitations on the allocation (only relevant for
462 * NUMA) that may require different handling.
464 constraint = constrained_alloc(zonelist, gfp_mask);
466 read_lock(&tasklist_lock);
468 switch (constraint) {
469 case CONSTRAINT_MEMORY_POLICY:
470 oom_kill_process(current, points,
471 "No available memory (MPOL_BIND)");
474 case CONSTRAINT_CPUSET:
475 oom_kill_process(current, points,
476 "No available memory in cpuset");
479 case CONSTRAINT_NONE:
480 if (sysctl_panic_on_oom)
481 panic("out of memory. panic_on_oom is selected\n");
484 * Rambo mode: Shoot down a process and hope it solves whatever
485 * issues we may have.
487 p = select_bad_process(&points);
489 if (PTR_ERR(p) == -1UL)
492 /* Found nothing?!?! Either we hang forever, or we panic. */
494 read_unlock(&tasklist_lock);
496 panic("Out of memory and no killable processes...\n");
499 if (oom_kill_process(p, points, "Out of memory"))
506 read_unlock(&tasklist_lock);
510 * Give "p" a good chance of killing itself before we
511 * retry to allocate memory unless "p" is current
513 if (!test_thread_flag(TIF_MEMDIE))
514 schedule_timeout_uninterruptible(1);