softlockup: fix NMI hangs due to lock race - 2.6.26-rc regression
[linux-2.6.git] / kernel / softlockup.c
1 /*
2  * Detect Soft Lockups
3  *
4  * started by Ingo Molnar, Copyright (C) 2005, 2006 Red Hat, Inc.
5  *
6  * this code detects soft lockups: incidents in where on a CPU
7  * the kernel does not reschedule for 10 seconds or more.
8  */
9 #include <linux/mm.h>
10 #include <linux/cpu.h>
11 #include <linux/nmi.h>
12 #include <linux/init.h>
13 #include <linux/delay.h>
14 #include <linux/freezer.h>
15 #include <linux/kthread.h>
16 #include <linux/notifier.h>
17 #include <linux/module.h>
18
19 #include <asm/irq_regs.h>
20
21 static DEFINE_SPINLOCK(print_lock);
22
23 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
24 static DEFINE_PER_CPU(unsigned long, print_timestamp);
25 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
26
27 static int __read_mostly did_panic;
28 unsigned long __read_mostly softlockup_thresh = 60;
29
30 static int
31 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
32 {
33         did_panic = 1;
34
35         return NOTIFY_DONE;
36 }
37
38 static struct notifier_block panic_block = {
39         .notifier_call = softlock_panic,
40 };
41
42 /*
43  * Returns seconds, approximately.  We don't need nanosecond
44  * resolution, and we don't need to waste time with a big divide when
45  * 2^30ns == 1.074s.
46  */
47 static unsigned long get_timestamp(int this_cpu)
48 {
49         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
50 }
51
52 static void __touch_softlockup_watchdog(void)
53 {
54         int this_cpu = raw_smp_processor_id();
55
56         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
57 }
58
59 void touch_softlockup_watchdog(void)
60 {
61         __raw_get_cpu_var(touch_timestamp) = 0;
62 }
63 EXPORT_SYMBOL(touch_softlockup_watchdog);
64
65 void touch_all_softlockup_watchdogs(void)
66 {
67         int cpu;
68
69         /* Cause each CPU to re-update its timestamp rather than complain */
70         for_each_online_cpu(cpu)
71                 per_cpu(touch_timestamp, cpu) = 0;
72 }
73 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
74
75 /*
76  * This callback runs from the timer interrupt, and checks
77  * whether the watchdog thread has hung or not:
78  */
79 void softlockup_tick(void)
80 {
81         int this_cpu = smp_processor_id();
82         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
83         unsigned long print_timestamp;
84         struct pt_regs *regs = get_irq_regs();
85         unsigned long now;
86
87         if (touch_timestamp == 0) {
88                 __touch_softlockup_watchdog();
89                 return;
90         }
91
92         print_timestamp = per_cpu(print_timestamp, this_cpu);
93
94         /* report at most once a second */
95         if ((print_timestamp >= touch_timestamp &&
96                         print_timestamp < (touch_timestamp + 1)) ||
97                         did_panic || !per_cpu(watchdog_task, this_cpu)) {
98                 return;
99         }
100
101         /* do not print during early bootup: */
102         if (unlikely(system_state != SYSTEM_RUNNING)) {
103                 __touch_softlockup_watchdog();
104                 return;
105         }
106
107         now = get_timestamp(this_cpu);
108
109         /* Wake up the high-prio watchdog task every second: */
110         if (now > (touch_timestamp + 1))
111                 wake_up_process(per_cpu(watchdog_task, this_cpu));
112
113         /* Warn about unreasonable delays: */
114         if (now <= (touch_timestamp + softlockup_thresh))
115                 return;
116
117         per_cpu(print_timestamp, this_cpu) = touch_timestamp;
118
119         spin_lock(&print_lock);
120         printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n",
121                         this_cpu, now - touch_timestamp,
122                         current->comm, task_pid_nr(current));
123         if (regs)
124                 show_regs(regs);
125         else
126                 dump_stack();
127         spin_unlock(&print_lock);
128 }
129
130 /*
131  * Have a reasonable limit on the number of tasks checked:
132  */
133 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
134
135 /*
136  * Zero means infinite timeout - no checking done:
137  */
138 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120;
139
140 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
141
142 /*
143  * Only do the hung-tasks check on one CPU:
144  */
145 static int check_cpu __read_mostly = -1;
146
147 static void check_hung_task(struct task_struct *t, unsigned long now)
148 {
149         unsigned long switch_count = t->nvcsw + t->nivcsw;
150
151         if (t->flags & PF_FROZEN)
152                 return;
153
154         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
155                 t->last_switch_count = switch_count;
156                 t->last_switch_timestamp = now;
157                 return;
158         }
159         if ((long)(now - t->last_switch_timestamp) <
160                                         sysctl_hung_task_timeout_secs)
161                 return;
162         if (sysctl_hung_task_warnings < 0)
163                 return;
164         sysctl_hung_task_warnings--;
165
166         /*
167          * Ok, the task did not get scheduled for more than 2 minutes,
168          * complain:
169          */
170         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
171                         "%ld seconds.\n", t->comm, t->pid,
172                         sysctl_hung_task_timeout_secs);
173         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
174                         " disables this message.\n");
175         sched_show_task(t);
176         __debug_show_held_locks(t);
177
178         t->last_switch_timestamp = now;
179         touch_nmi_watchdog();
180 }
181
182 /*
183  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
184  * a really long time (120 seconds). If that happens, print out
185  * a warning.
186  */
187 static void check_hung_uninterruptible_tasks(int this_cpu)
188 {
189         int max_count = sysctl_hung_task_check_count;
190         unsigned long now = get_timestamp(this_cpu);
191         struct task_struct *g, *t;
192
193         /*
194          * If the system crashed already then all bets are off,
195          * do not report extra hung tasks:
196          */
197         if ((tainted & TAINT_DIE) || did_panic)
198                 return;
199
200         read_lock(&tasklist_lock);
201         do_each_thread(g, t) {
202                 if (!--max_count)
203                         goto unlock;
204                 if (t->state & TASK_UNINTERRUPTIBLE)
205                         check_hung_task(t, now);
206         } while_each_thread(g, t);
207  unlock:
208         read_unlock(&tasklist_lock);
209 }
210
211 /*
212  * The watchdog thread - runs every second and touches the timestamp.
213  */
214 static int watchdog(void *__bind_cpu)
215 {
216         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
217         int this_cpu = (long)__bind_cpu;
218
219         sched_setscheduler(current, SCHED_FIFO, &param);
220
221         /* initialize timestamp */
222         __touch_softlockup_watchdog();
223
224         set_current_state(TASK_INTERRUPTIBLE);
225         /*
226          * Run briefly once per second to reset the softlockup timestamp.
227          * If this gets delayed for more than 60 seconds then the
228          * debug-printout triggers in softlockup_tick().
229          */
230         while (!kthread_should_stop()) {
231                 __touch_softlockup_watchdog();
232                 schedule();
233
234                 if (kthread_should_stop())
235                         break;
236
237                 if (this_cpu == check_cpu) {
238                         if (sysctl_hung_task_timeout_secs)
239                                 check_hung_uninterruptible_tasks(this_cpu);
240                 }
241
242                 set_current_state(TASK_INTERRUPTIBLE);
243         }
244         __set_current_state(TASK_RUNNING);
245
246         return 0;
247 }
248
249 /*
250  * Create/destroy watchdog threads as CPUs come and go:
251  */
252 static int __cpuinit
253 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
254 {
255         int hotcpu = (unsigned long)hcpu;
256         struct task_struct *p;
257
258         switch (action) {
259         case CPU_UP_PREPARE:
260         case CPU_UP_PREPARE_FROZEN:
261                 BUG_ON(per_cpu(watchdog_task, hotcpu));
262                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
263                 if (IS_ERR(p)) {
264                         printk(KERN_ERR "watchdog for %i failed\n", hotcpu);
265                         return NOTIFY_BAD;
266                 }
267                 per_cpu(touch_timestamp, hotcpu) = 0;
268                 per_cpu(watchdog_task, hotcpu) = p;
269                 kthread_bind(p, hotcpu);
270                 break;
271         case CPU_ONLINE:
272         case CPU_ONLINE_FROZEN:
273                 check_cpu = any_online_cpu(cpu_online_map);
274                 wake_up_process(per_cpu(watchdog_task, hotcpu));
275                 break;
276 #ifdef CONFIG_HOTPLUG_CPU
277         case CPU_DOWN_PREPARE:
278         case CPU_DOWN_PREPARE_FROZEN:
279                 if (hotcpu == check_cpu) {
280                         cpumask_t temp_cpu_online_map = cpu_online_map;
281
282                         cpu_clear(hotcpu, temp_cpu_online_map);
283                         check_cpu = any_online_cpu(temp_cpu_online_map);
284                 }
285                 break;
286
287         case CPU_UP_CANCELED:
288         case CPU_UP_CANCELED_FROZEN:
289                 if (!per_cpu(watchdog_task, hotcpu))
290                         break;
291                 /* Unbind so it can run.  Fall thru. */
292                 kthread_bind(per_cpu(watchdog_task, hotcpu),
293                              any_online_cpu(cpu_online_map));
294         case CPU_DEAD:
295         case CPU_DEAD_FROZEN:
296                 p = per_cpu(watchdog_task, hotcpu);
297                 per_cpu(watchdog_task, hotcpu) = NULL;
298                 kthread_stop(p);
299                 break;
300 #endif /* CONFIG_HOTPLUG_CPU */
301         }
302         return NOTIFY_OK;
303 }
304
305 static struct notifier_block __cpuinitdata cpu_nfb = {
306         .notifier_call = cpu_callback
307 };
308
309 __init void spawn_softlockup_task(void)
310 {
311         void *cpu = (void *)(long)smp_processor_id();
312         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
313
314         BUG_ON(err == NOTIFY_BAD);
315         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
316         register_cpu_notifier(&cpu_nfb);
317
318         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
319 }