Commit | Line | Data |
---|---|---|
e162b39a MSB |
1 | /* |
2 | * Detect Hung Task | |
3 | * | |
4 | * kernel/hung_task.c - kernel thread for detecting tasks stuck in D state | |
5 | * | |
6 | */ | |
7 | ||
8 | #include <linux/mm.h> | |
9 | #include <linux/cpu.h> | |
10 | #include <linux/nmi.h> | |
11 | #include <linux/init.h> | |
12 | #include <linux/delay.h> | |
13 | #include <linux/freezer.h> | |
14 | #include <linux/kthread.h> | |
15 | #include <linux/lockdep.h> | |
9984de1a | 16 | #include <linux/export.h> |
e162b39a MSB |
17 | #include <linux/sysctl.h> |
18 | ||
19 | /* | |
ce9dbe24 | 20 | * The number of tasks checked: |
e162b39a | 21 | */ |
ce9dbe24 MSB |
22 | unsigned long __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT; |
23 | ||
24 | /* | |
25 | * Limit number of tasks checked in a batch. | |
26 | * | |
27 | * This value controls the preemptibility of khungtaskd since preemption | |
28 | * is disabled during the critical section. It also controls the size of | |
29 | * the RCU grace period. So it needs to be upper-bound. | |
30 | */ | |
31 | #define HUNG_TASK_BATCHING 1024 | |
e162b39a MSB |
32 | |
33 | /* | |
34 | * Zero means infinite timeout - no checking done: | |
35 | */ | |
e11feaa1 | 36 | unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT; |
e162b39a MSB |
37 | |
38 | unsigned long __read_mostly sysctl_hung_task_warnings = 10; | |
39 | ||
40 | static int __read_mostly did_panic; | |
41 | ||
42 | static struct task_struct *watchdog_task; | |
43 | ||
44 | /* | |
45 | * Should we panic (and reboot, if panic_timeout= is set) when a | |
46 | * hung task is detected: | |
47 | */ | |
48 | unsigned int __read_mostly sysctl_hung_task_panic = | |
49 | CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE; | |
50 | ||
51 | static int __init hung_task_panic_setup(char *str) | |
52 | { | |
53 | sysctl_hung_task_panic = simple_strtoul(str, NULL, 0); | |
54 | ||
55 | return 1; | |
56 | } | |
57 | __setup("hung_task_panic=", hung_task_panic_setup); | |
58 | ||
59 | static int | |
60 | hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr) | |
61 | { | |
62 | did_panic = 1; | |
63 | ||
64 | return NOTIFY_DONE; | |
65 | } | |
66 | ||
67 | static struct notifier_block panic_block = { | |
68 | .notifier_call = hung_task_panic, | |
69 | }; | |
70 | ||
17406b82 | 71 | static void check_hung_task(struct task_struct *t, unsigned long timeout) |
e162b39a MSB |
72 | { |
73 | unsigned long switch_count = t->nvcsw + t->nivcsw; | |
74 | ||
cf2592f5 FW |
75 | /* |
76 | * Ensure the task is not frozen. | |
f9fab10b | 77 | * Also, skip vfork and any other user process that freezer should skip. |
cf2592f5 | 78 | */ |
f9fab10b MSB |
79 | if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP))) |
80 | return; | |
81 | ||
82 | /* | |
83 | * When a freshly created task is scheduled once, changes its state to | |
84 | * TASK_UNINTERRUPTIBLE without having ever been switched out once, it | |
85 | * musn't be checked. | |
86 | */ | |
87 | if (unlikely(!switch_count)) | |
e162b39a MSB |
88 | return; |
89 | ||
17406b82 | 90 | if (switch_count != t->last_switch_count) { |
e162b39a | 91 | t->last_switch_count = switch_count; |
e162b39a MSB |
92 | return; |
93 | } | |
e162b39a MSB |
94 | if (!sysctl_hung_task_warnings) |
95 | return; | |
96 | sysctl_hung_task_warnings--; | |
97 | ||
98 | /* | |
99 | * Ok, the task did not get scheduled for more than 2 minutes, | |
100 | * complain: | |
101 | */ | |
102 | printk(KERN_ERR "INFO: task %s:%d blocked for more than " | |
603a148f | 103 | "%ld seconds.\n", t->comm, t->pid, timeout); |
e162b39a MSB |
104 | printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\"" |
105 | " disables this message.\n"); | |
106 | sched_show_task(t); | |
f1b499f0 | 107 | debug_show_held_locks(t); |
e162b39a | 108 | |
e162b39a MSB |
109 | touch_nmi_watchdog(); |
110 | ||
111 | if (sysctl_hung_task_panic) | |
112 | panic("hung_task: blocked tasks"); | |
113 | } | |
114 | ||
ce9dbe24 MSB |
115 | /* |
116 | * To avoid extending the RCU grace period for an unbounded amount of time, | |
117 | * periodically exit the critical section and enter a new one. | |
118 | * | |
119 | * For preemptible RCU it is sufficient to call rcu_read_unlock in order | |
6a103b0d | 120 | * to exit the grace period. For classic RCU, a reschedule is required. |
ce9dbe24 | 121 | */ |
6027ce49 | 122 | static bool rcu_lock_break(struct task_struct *g, struct task_struct *t) |
ce9dbe24 | 123 | { |
6027ce49 ON |
124 | bool can_cont; |
125 | ||
ce9dbe24 MSB |
126 | get_task_struct(g); |
127 | get_task_struct(t); | |
128 | rcu_read_unlock(); | |
129 | cond_resched(); | |
130 | rcu_read_lock(); | |
6027ce49 | 131 | can_cont = pid_alive(g) && pid_alive(t); |
ce9dbe24 MSB |
132 | put_task_struct(t); |
133 | put_task_struct(g); | |
6027ce49 ON |
134 | |
135 | return can_cont; | |
ce9dbe24 MSB |
136 | } |
137 | ||
e162b39a MSB |
138 | /* |
139 | * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for | |
140 | * a really long time (120 seconds). If that happens, print out | |
141 | * a warning. | |
142 | */ | |
603a148f | 143 | static void check_hung_uninterruptible_tasks(unsigned long timeout) |
e162b39a MSB |
144 | { |
145 | int max_count = sysctl_hung_task_check_count; | |
ce9dbe24 | 146 | int batch_count = HUNG_TASK_BATCHING; |
e162b39a MSB |
147 | struct task_struct *g, *t; |
148 | ||
149 | /* | |
150 | * If the system crashed already then all bets are off, | |
151 | * do not report extra hung tasks: | |
152 | */ | |
153 | if (test_taint(TAINT_DIE) || did_panic) | |
154 | return; | |
155 | ||
94be52dc | 156 | rcu_read_lock(); |
e162b39a | 157 | do_each_thread(g, t) { |
e5af0226 | 158 | if (!max_count--) |
e162b39a | 159 | goto unlock; |
ce9dbe24 MSB |
160 | if (!--batch_count) { |
161 | batch_count = HUNG_TASK_BATCHING; | |
6027ce49 | 162 | if (!rcu_lock_break(g, t)) |
ce9dbe24 MSB |
163 | goto unlock; |
164 | } | |
e162b39a MSB |
165 | /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */ |
166 | if (t->state == TASK_UNINTERRUPTIBLE) | |
17406b82 | 167 | check_hung_task(t, timeout); |
e162b39a MSB |
168 | } while_each_thread(g, t); |
169 | unlock: | |
94be52dc | 170 | rcu_read_unlock(); |
e162b39a MSB |
171 | } |
172 | ||
17406b82 | 173 | static unsigned long timeout_jiffies(unsigned long timeout) |
e162b39a MSB |
174 | { |
175 | /* timeout of 0 will disable the watchdog */ | |
17406b82 | 176 | return timeout ? timeout * HZ : MAX_SCHEDULE_TIMEOUT; |
e162b39a MSB |
177 | } |
178 | ||
179 | /* | |
180 | * Process updating of timeout sysctl | |
181 | */ | |
182 | int proc_dohung_task_timeout_secs(struct ctl_table *table, int write, | |
8d65af78 | 183 | void __user *buffer, |
e162b39a MSB |
184 | size_t *lenp, loff_t *ppos) |
185 | { | |
186 | int ret; | |
187 | ||
8d65af78 | 188 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
e162b39a MSB |
189 | |
190 | if (ret || !write) | |
191 | goto out; | |
192 | ||
e162b39a MSB |
193 | wake_up_process(watchdog_task); |
194 | ||
195 | out: | |
196 | return ret; | |
197 | } | |
198 | ||
199 | /* | |
200 | * kthread which checks for tasks stuck in D state | |
201 | */ | |
202 | static int watchdog(void *dummy) | |
203 | { | |
204 | set_user_nice(current, 0); | |
e162b39a MSB |
205 | |
206 | for ( ; ; ) { | |
17406b82 | 207 | unsigned long timeout = sysctl_hung_task_timeout_secs; |
603a148f | 208 | |
17406b82 MSB |
209 | while (schedule_timeout_interruptible(timeout_jiffies(timeout))) |
210 | timeout = sysctl_hung_task_timeout_secs; | |
603a148f | 211 | |
17406b82 | 212 | check_hung_uninterruptible_tasks(timeout); |
e162b39a MSB |
213 | } |
214 | ||
215 | return 0; | |
216 | } | |
217 | ||
218 | static int __init hung_task_init(void) | |
219 | { | |
220 | atomic_notifier_chain_register(&panic_notifier_list, &panic_block); | |
221 | watchdog_task = kthread_run(watchdog, NULL, "khungtaskd"); | |
222 | ||
223 | return 0; | |
224 | } | |
225 | ||
226 | module_init(hung_task_init); |