| 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/lockdep.h> |
| 17 | #include <linux/notifier.h> |
| 18 | #include <linux/module.h> |
| 19 | #include <linux/sysctl.h> |
| 20 | |
| 21 | #include <asm/irq_regs.h> |
| 22 | |
| 23 | static DEFINE_SPINLOCK(print_lock); |
| 24 | |
| 25 | static DEFINE_PER_CPU(unsigned long, softlockup_touch_ts); /* touch timestamp */ |
| 26 | static DEFINE_PER_CPU(unsigned long, softlockup_print_ts); /* print timestamp */ |
| 27 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
| 28 | static DEFINE_PER_CPU(bool, softlock_touch_sync); |
| 29 | |
| 30 | static int __read_mostly did_panic; |
| 31 | int __read_mostly softlockup_thresh = 60; |
| 32 | |
| 33 | /* |
| 34 | * Should we panic (and reboot, if panic_timeout= is set) when a |
| 35 | * soft-lockup occurs: |
| 36 | */ |
| 37 | unsigned int __read_mostly softlockup_panic = |
| 38 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; |
| 39 | |
| 40 | static int __init softlockup_panic_setup(char *str) |
| 41 | { |
| 42 | softlockup_panic = simple_strtoul(str, NULL, 0); |
| 43 | |
| 44 | return 1; |
| 45 | } |
| 46 | __setup("softlockup_panic=", softlockup_panic_setup); |
| 47 | |
| 48 | static int |
| 49 | softlock_panic(struct notifier_block *this, unsigned long event, void *ptr) |
| 50 | { |
| 51 | did_panic = 1; |
| 52 | |
| 53 | return NOTIFY_DONE; |
| 54 | } |
| 55 | |
| 56 | static struct notifier_block panic_block = { |
| 57 | .notifier_call = softlock_panic, |
| 58 | }; |
| 59 | |
| 60 | /* |
| 61 | * Returns seconds, approximately. We don't need nanosecond |
| 62 | * resolution, and we don't need to waste time with a big divide when |
| 63 | * 2^30ns == 1.074s. |
| 64 | */ |
| 65 | static unsigned long get_timestamp(int this_cpu) |
| 66 | { |
| 67 | return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */ |
| 68 | } |
| 69 | |
| 70 | static void __touch_softlockup_watchdog(void) |
| 71 | { |
| 72 | int this_cpu = raw_smp_processor_id(); |
| 73 | |
| 74 | __raw_get_cpu_var(softlockup_touch_ts) = get_timestamp(this_cpu); |
| 75 | } |
| 76 | |
| 77 | void touch_softlockup_watchdog(void) |
| 78 | { |
| 79 | __raw_get_cpu_var(softlockup_touch_ts) = 0; |
| 80 | } |
| 81 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
| 82 | |
| 83 | void touch_softlockup_watchdog_sync(void) |
| 84 | { |
| 85 | __raw_get_cpu_var(softlock_touch_sync) = true; |
| 86 | __raw_get_cpu_var(softlockup_touch_ts) = 0; |
| 87 | } |
| 88 | |
| 89 | void touch_all_softlockup_watchdogs(void) |
| 90 | { |
| 91 | int cpu; |
| 92 | |
| 93 | /* Cause each CPU to re-update its timestamp rather than complain */ |
| 94 | for_each_online_cpu(cpu) |
| 95 | per_cpu(softlockup_touch_ts, cpu) = 0; |
| 96 | } |
| 97 | EXPORT_SYMBOL(touch_all_softlockup_watchdogs); |
| 98 | |
| 99 | int proc_dosoftlockup_thresh(struct ctl_table *table, int write, |
| 100 | void __user *buffer, |
| 101 | size_t *lenp, loff_t *ppos) |
| 102 | { |
| 103 | touch_all_softlockup_watchdogs(); |
| 104 | return proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 105 | } |
| 106 | |
| 107 | /* |
| 108 | * This callback runs from the timer interrupt, and checks |
| 109 | * whether the watchdog thread has hung or not: |
| 110 | */ |
| 111 | void softlockup_tick(void) |
| 112 | { |
| 113 | int this_cpu = smp_processor_id(); |
| 114 | unsigned long touch_ts = per_cpu(softlockup_touch_ts, this_cpu); |
| 115 | unsigned long print_ts; |
| 116 | struct pt_regs *regs = get_irq_regs(); |
| 117 | unsigned long now; |
| 118 | |
| 119 | /* Is detection switched off? */ |
| 120 | if (!per_cpu(softlockup_watchdog, this_cpu) || softlockup_thresh <= 0) { |
| 121 | /* Be sure we don't false trigger if switched back on */ |
| 122 | if (touch_ts) |
| 123 | per_cpu(softlockup_touch_ts, this_cpu) = 0; |
| 124 | return; |
| 125 | } |
| 126 | |
| 127 | if (touch_ts == 0) { |
| 128 | if (unlikely(per_cpu(softlock_touch_sync, this_cpu))) { |
| 129 | /* |
| 130 | * If the time stamp was touched atomically |
| 131 | * make sure the scheduler tick is up to date. |
| 132 | */ |
| 133 | per_cpu(softlock_touch_sync, this_cpu) = false; |
| 134 | sched_clock_tick(); |
| 135 | } |
| 136 | __touch_softlockup_watchdog(); |
| 137 | return; |
| 138 | } |
| 139 | |
| 140 | print_ts = per_cpu(softlockup_print_ts, this_cpu); |
| 141 | |
| 142 | /* report at most once a second */ |
| 143 | if (print_ts == touch_ts || did_panic) |
| 144 | return; |
| 145 | |
| 146 | /* do not print during early bootup: */ |
| 147 | if (unlikely(system_state != SYSTEM_RUNNING)) { |
| 148 | __touch_softlockup_watchdog(); |
| 149 | return; |
| 150 | } |
| 151 | |
| 152 | now = get_timestamp(this_cpu); |
| 153 | |
| 154 | /* |
| 155 | * Wake up the high-prio watchdog task twice per |
| 156 | * threshold timespan. |
| 157 | */ |
| 158 | if (time_after(now - softlockup_thresh/2, touch_ts)) |
| 159 | wake_up_process(per_cpu(softlockup_watchdog, this_cpu)); |
| 160 | |
| 161 | /* Warn about unreasonable delays: */ |
| 162 | if (time_before_eq(now - softlockup_thresh, touch_ts)) |
| 163 | return; |
| 164 | |
| 165 | per_cpu(softlockup_print_ts, this_cpu) = touch_ts; |
| 166 | |
| 167 | spin_lock(&print_lock); |
| 168 | printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n", |
| 169 | this_cpu, now - touch_ts, |
| 170 | current->comm, task_pid_nr(current)); |
| 171 | print_modules(); |
| 172 | print_irqtrace_events(current); |
| 173 | if (regs) |
| 174 | show_regs(regs); |
| 175 | else |
| 176 | dump_stack(); |
| 177 | spin_unlock(&print_lock); |
| 178 | |
| 179 | if (softlockup_panic) |
| 180 | panic("softlockup: hung tasks"); |
| 181 | } |
| 182 | |
| 183 | /* |
| 184 | * The watchdog thread - runs every second and touches the timestamp. |
| 185 | */ |
| 186 | static int watchdog(void *__bind_cpu) |
| 187 | { |
| 188 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; |
| 189 | |
| 190 | sched_setscheduler(current, SCHED_FIFO, ¶m); |
| 191 | |
| 192 | /* initialize timestamp */ |
| 193 | __touch_softlockup_watchdog(); |
| 194 | |
| 195 | set_current_state(TASK_INTERRUPTIBLE); |
| 196 | /* |
| 197 | * Run briefly once per second to reset the softlockup timestamp. |
| 198 | * If this gets delayed for more than 60 seconds then the |
| 199 | * debug-printout triggers in softlockup_tick(). |
| 200 | */ |
| 201 | while (!kthread_should_stop()) { |
| 202 | __touch_softlockup_watchdog(); |
| 203 | schedule(); |
| 204 | |
| 205 | if (kthread_should_stop()) |
| 206 | break; |
| 207 | |
| 208 | set_current_state(TASK_INTERRUPTIBLE); |
| 209 | } |
| 210 | __set_current_state(TASK_RUNNING); |
| 211 | |
| 212 | return 0; |
| 213 | } |
| 214 | |
| 215 | /* |
| 216 | * Create/destroy watchdog threads as CPUs come and go: |
| 217 | */ |
| 218 | static int __cpuinit |
| 219 | cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 220 | { |
| 221 | int hotcpu = (unsigned long)hcpu; |
| 222 | struct task_struct *p; |
| 223 | |
| 224 | switch (action) { |
| 225 | case CPU_UP_PREPARE: |
| 226 | case CPU_UP_PREPARE_FROZEN: |
| 227 | BUG_ON(per_cpu(softlockup_watchdog, hotcpu)); |
| 228 | p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu); |
| 229 | if (IS_ERR(p)) { |
| 230 | printk(KERN_ERR "watchdog for %i failed\n", hotcpu); |
| 231 | return NOTIFY_BAD; |
| 232 | } |
| 233 | per_cpu(softlockup_touch_ts, hotcpu) = 0; |
| 234 | per_cpu(softlockup_watchdog, hotcpu) = p; |
| 235 | kthread_bind(p, hotcpu); |
| 236 | break; |
| 237 | case CPU_ONLINE: |
| 238 | case CPU_ONLINE_FROZEN: |
| 239 | wake_up_process(per_cpu(softlockup_watchdog, hotcpu)); |
| 240 | break; |
| 241 | #ifdef CONFIG_HOTPLUG_CPU |
| 242 | case CPU_UP_CANCELED: |
| 243 | case CPU_UP_CANCELED_FROZEN: |
| 244 | if (!per_cpu(softlockup_watchdog, hotcpu)) |
| 245 | break; |
| 246 | /* Unbind so it can run. Fall thru. */ |
| 247 | kthread_bind(per_cpu(softlockup_watchdog, hotcpu), |
| 248 | cpumask_any(cpu_online_mask)); |
| 249 | case CPU_DEAD: |
| 250 | case CPU_DEAD_FROZEN: |
| 251 | p = per_cpu(softlockup_watchdog, hotcpu); |
| 252 | per_cpu(softlockup_watchdog, hotcpu) = NULL; |
| 253 | kthread_stop(p); |
| 254 | break; |
| 255 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 256 | } |
| 257 | return NOTIFY_OK; |
| 258 | } |
| 259 | |
| 260 | static struct notifier_block __cpuinitdata cpu_nfb = { |
| 261 | .notifier_call = cpu_callback |
| 262 | }; |
| 263 | |
| 264 | static int __initdata nosoftlockup; |
| 265 | |
| 266 | static int __init nosoftlockup_setup(char *str) |
| 267 | { |
| 268 | nosoftlockup = 1; |
| 269 | return 1; |
| 270 | } |
| 271 | __setup("nosoftlockup", nosoftlockup_setup); |
| 272 | |
| 273 | static int __init spawn_softlockup_task(void) |
| 274 | { |
| 275 | void *cpu = (void *)(long)smp_processor_id(); |
| 276 | int err; |
| 277 | |
| 278 | if (nosoftlockup) |
| 279 | return 0; |
| 280 | |
| 281 | err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu); |
| 282 | if (err == NOTIFY_BAD) { |
| 283 | BUG(); |
| 284 | return 1; |
| 285 | } |
| 286 | cpu_callback(&cpu_nfb, CPU_ONLINE, cpu); |
| 287 | register_cpu_notifier(&cpu_nfb); |
| 288 | |
| 289 | atomic_notifier_chain_register(&panic_notifier_list, &panic_block); |
| 290 | |
| 291 | return 0; |
| 292 | } |
| 293 | early_initcall(spawn_softlockup_task); |