Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Detect hard and soft lockups on a system |
| 3 | * |
| 4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. |
| 5 | * |
| 6 | * this code detects hard lockups: incidents in where on a CPU |
| 7 | * the kernel does not respond to anything except NMI. |
| 8 | * |
| 9 | * Note: Most of this code is borrowed heavily from softlockup.c, |
| 10 | * so thanks to Ingo for the initial implementation. |
| 11 | * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks |
| 12 | * to those contributors as well. |
| 13 | */ |
| 14 | |
| 15 | #include <linux/mm.h> |
| 16 | #include <linux/cpu.h> |
| 17 | #include <linux/nmi.h> |
| 18 | #include <linux/init.h> |
| 19 | #include <linux/delay.h> |
| 20 | #include <linux/freezer.h> |
| 21 | #include <linux/kthread.h> |
| 22 | #include <linux/lockdep.h> |
| 23 | #include <linux/notifier.h> |
| 24 | #include <linux/module.h> |
| 25 | #include <linux/sysctl.h> |
| 26 | |
| 27 | #include <asm/irq_regs.h> |
| 28 | #include <linux/perf_event.h> |
| 29 | |
| 30 | int watchdog_enabled; |
| 31 | int __read_mostly softlockup_thresh = 60; |
| 32 | |
| 33 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); |
| 34 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
| 35 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); |
| 36 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 37 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 38 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | cafcd80 | 2010-05-14 11:11:21 -0400 | [diff] [blame] | 39 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
| 40 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 41 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
| 42 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
| 43 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); |
| 44 | #endif |
| 45 | |
David Daney | 433039e | 2010-11-05 16:17:39 -0700 | [diff] [blame] | 46 | static int no_watchdog; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 47 | |
| 48 | |
| 49 | /* boot commands */ |
| 50 | /* |
| 51 | * Should we panic when a soft-lockup or hard-lockup occurs: |
| 52 | */ |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 53 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 54 | static int hardlockup_panic; |
| 55 | |
| 56 | static int __init hardlockup_panic_setup(char *str) |
| 57 | { |
| 58 | if (!strncmp(str, "panic", 5)) |
| 59 | hardlockup_panic = 1; |
Don Zickus | 5dc3055 | 2010-11-29 17:07:17 -0500 | [diff] [blame] | 60 | else if (!strncmp(str, "0", 1)) |
| 61 | no_watchdog = 1; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 62 | return 1; |
| 63 | } |
| 64 | __setup("nmi_watchdog=", hardlockup_panic_setup); |
| 65 | #endif |
| 66 | |
| 67 | unsigned int __read_mostly softlockup_panic = |
| 68 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; |
| 69 | |
| 70 | static int __init softlockup_panic_setup(char *str) |
| 71 | { |
| 72 | softlockup_panic = simple_strtoul(str, NULL, 0); |
| 73 | |
| 74 | return 1; |
| 75 | } |
| 76 | __setup("softlockup_panic=", softlockup_panic_setup); |
| 77 | |
| 78 | static int __init nowatchdog_setup(char *str) |
| 79 | { |
| 80 | no_watchdog = 1; |
| 81 | return 1; |
| 82 | } |
| 83 | __setup("nowatchdog", nowatchdog_setup); |
| 84 | |
| 85 | /* deprecated */ |
| 86 | static int __init nosoftlockup_setup(char *str) |
| 87 | { |
| 88 | no_watchdog = 1; |
| 89 | return 1; |
| 90 | } |
| 91 | __setup("nosoftlockup", nosoftlockup_setup); |
| 92 | /* */ |
| 93 | |
| 94 | |
| 95 | /* |
| 96 | * Returns seconds, approximately. We don't need nanosecond |
| 97 | * resolution, and we don't need to waste time with a big divide when |
| 98 | * 2^30ns == 1.074s. |
| 99 | */ |
| 100 | static unsigned long get_timestamp(int this_cpu) |
| 101 | { |
| 102 | return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */ |
| 103 | } |
| 104 | |
| 105 | static unsigned long get_sample_period(void) |
| 106 | { |
| 107 | /* |
| 108 | * convert softlockup_thresh from seconds to ns |
| 109 | * the divide by 5 is to give hrtimer 5 chances to |
| 110 | * increment before the hardlockup detector generates |
| 111 | * a warning |
| 112 | */ |
| 113 | return softlockup_thresh / 5 * NSEC_PER_SEC; |
| 114 | } |
| 115 | |
| 116 | /* Commands for resetting the watchdog */ |
| 117 | static void __touch_watchdog(void) |
| 118 | { |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 119 | int this_cpu = smp_processor_id(); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 120 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 121 | __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu)); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 122 | } |
| 123 | |
Don Zickus | 332fbdb | 2010-05-07 17:11:45 -0400 | [diff] [blame] | 124 | void touch_softlockup_watchdog(void) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 125 | { |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 126 | __this_cpu_write(watchdog_touch_ts, 0); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 127 | } |
Ingo Molnar | 0167c78 | 2010-05-13 08:53:33 +0200 | [diff] [blame] | 128 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 129 | |
Don Zickus | 332fbdb | 2010-05-07 17:11:45 -0400 | [diff] [blame] | 130 | void touch_all_softlockup_watchdogs(void) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 131 | { |
| 132 | int cpu; |
| 133 | |
| 134 | /* |
| 135 | * this is done lockless |
| 136 | * do we care if a 0 races with a timestamp? |
| 137 | * all it means is the softlock check starts one cycle later |
| 138 | */ |
| 139 | for_each_online_cpu(cpu) |
| 140 | per_cpu(watchdog_touch_ts, cpu) = 0; |
| 141 | } |
| 142 | |
Don Zickus | cafcd80 | 2010-05-14 11:11:21 -0400 | [diff] [blame] | 143 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 144 | void touch_nmi_watchdog(void) |
| 145 | { |
Don Zickus | 68d3f1d | 2010-08-31 23:00:07 -0400 | [diff] [blame] | 146 | if (watchdog_enabled) { |
| 147 | unsigned cpu; |
| 148 | |
| 149 | for_each_present_cpu(cpu) { |
| 150 | if (per_cpu(watchdog_nmi_touch, cpu) != true) |
| 151 | per_cpu(watchdog_nmi_touch, cpu) = true; |
| 152 | } |
| 153 | } |
Don Zickus | 332fbdb | 2010-05-07 17:11:45 -0400 | [diff] [blame] | 154 | touch_softlockup_watchdog(); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 155 | } |
| 156 | EXPORT_SYMBOL(touch_nmi_watchdog); |
| 157 | |
Don Zickus | cafcd80 | 2010-05-14 11:11:21 -0400 | [diff] [blame] | 158 | #endif |
| 159 | |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 160 | void touch_softlockup_watchdog_sync(void) |
| 161 | { |
| 162 | __raw_get_cpu_var(softlockup_touch_sync) = true; |
| 163 | __raw_get_cpu_var(watchdog_touch_ts) = 0; |
| 164 | } |
| 165 | |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 166 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 167 | /* watchdog detector functions */ |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 168 | static int is_hardlockup(void) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 169 | { |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 170 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 171 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 172 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 173 | return 1; |
| 174 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 175 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 176 | return 0; |
| 177 | } |
| 178 | #endif |
| 179 | |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 180 | static int is_softlockup(unsigned long touch_ts) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 181 | { |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 182 | unsigned long now = get_timestamp(smp_processor_id()); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 183 | |
| 184 | /* Warn about unreasonable delays: */ |
| 185 | if (time_after(now, touch_ts + softlockup_thresh)) |
| 186 | return now - touch_ts; |
| 187 | |
| 188 | return 0; |
| 189 | } |
| 190 | |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 191 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 192 | static struct perf_event_attr wd_hw_attr = { |
| 193 | .type = PERF_TYPE_HARDWARE, |
| 194 | .config = PERF_COUNT_HW_CPU_CYCLES, |
| 195 | .size = sizeof(struct perf_event_attr), |
| 196 | .pinned = 1, |
| 197 | .disabled = 1, |
| 198 | }; |
| 199 | |
| 200 | /* Callback function for perf event subsystem */ |
Lin Ming | 277b199 | 2010-08-20 11:03:51 +0800 | [diff] [blame] | 201 | static void watchdog_overflow_callback(struct perf_event *event, int nmi, |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 202 | struct perf_sample_data *data, |
| 203 | struct pt_regs *regs) |
| 204 | { |
Peter Zijlstra | c6db67c | 2010-08-20 11:49:15 +0200 | [diff] [blame] | 205 | /* Ensure the watchdog never gets throttled */ |
| 206 | event->hw.interrupts = 0; |
| 207 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 208 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
| 209 | __this_cpu_write(watchdog_nmi_touch, false); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 210 | return; |
| 211 | } |
| 212 | |
| 213 | /* check for a hardlockup |
| 214 | * This is done by making sure our timer interrupt |
| 215 | * is incrementing. The timer interrupt should have |
| 216 | * fired multiple times before we overflow'd. If it hasn't |
| 217 | * then this is a good indication the cpu is stuck |
| 218 | */ |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 219 | if (is_hardlockup()) { |
| 220 | int this_cpu = smp_processor_id(); |
| 221 | |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 222 | /* only print hardlockups once */ |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 223 | if (__this_cpu_read(hard_watchdog_warn) == true) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 224 | return; |
| 225 | |
| 226 | if (hardlockup_panic) |
| 227 | panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu); |
| 228 | else |
| 229 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu); |
| 230 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 231 | __this_cpu_write(hard_watchdog_warn, true); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 232 | return; |
| 233 | } |
| 234 | |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 235 | __this_cpu_write(hard_watchdog_warn, false); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 236 | return; |
| 237 | } |
| 238 | static void watchdog_interrupt_count(void) |
| 239 | { |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 240 | __this_cpu_inc(hrtimer_interrupts); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 241 | } |
| 242 | #else |
| 243 | static inline void watchdog_interrupt_count(void) { return; } |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 244 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 245 | |
| 246 | /* watchdog kicker functions */ |
| 247 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) |
| 248 | { |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 249 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 250 | struct pt_regs *regs = get_irq_regs(); |
| 251 | int duration; |
| 252 | |
| 253 | /* kick the hardlockup detector */ |
| 254 | watchdog_interrupt_count(); |
| 255 | |
| 256 | /* kick the softlockup detector */ |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 257 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 258 | |
| 259 | /* .. and repeat */ |
| 260 | hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period())); |
| 261 | |
| 262 | if (touch_ts == 0) { |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 263 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 264 | /* |
| 265 | * If the time stamp was touched atomically |
| 266 | * make sure the scheduler tick is up to date. |
| 267 | */ |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 268 | __this_cpu_write(softlockup_touch_sync, false); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 269 | sched_clock_tick(); |
| 270 | } |
| 271 | __touch_watchdog(); |
| 272 | return HRTIMER_RESTART; |
| 273 | } |
| 274 | |
| 275 | /* check for a softlockup |
| 276 | * This is done by making sure a high priority task is |
| 277 | * being scheduled. The task touches the watchdog to |
| 278 | * indicate it is getting cpu time. If it hasn't then |
| 279 | * this is a good indication some task is hogging the cpu |
| 280 | */ |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 281 | duration = is_softlockup(touch_ts); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 282 | if (unlikely(duration)) { |
| 283 | /* only warn once */ |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 284 | if (__this_cpu_read(soft_watchdog_warn) == true) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 285 | return HRTIMER_RESTART; |
| 286 | |
| 287 | printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 288 | smp_processor_id(), duration, |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 289 | current->comm, task_pid_nr(current)); |
| 290 | print_modules(); |
| 291 | print_irqtrace_events(current); |
| 292 | if (regs) |
| 293 | show_regs(regs); |
| 294 | else |
| 295 | dump_stack(); |
| 296 | |
| 297 | if (softlockup_panic) |
| 298 | panic("softlockup: hung tasks"); |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 299 | __this_cpu_write(soft_watchdog_warn, true); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 300 | } else |
Christoph Lameter | 909ea96 | 2010-12-08 16:22:55 +0100 | [diff] [blame] | 301 | __this_cpu_write(soft_watchdog_warn, false); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 302 | |
| 303 | return HRTIMER_RESTART; |
| 304 | } |
| 305 | |
| 306 | |
| 307 | /* |
| 308 | * The watchdog thread - touches the timestamp. |
| 309 | */ |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 310 | static int watchdog(void *unused) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 311 | { |
KOSAKI Motohiro | fe7de49 | 2010-10-20 16:01:12 -0700 | [diff] [blame] | 312 | static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 313 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 314 | |
| 315 | sched_setscheduler(current, SCHED_FIFO, ¶m); |
| 316 | |
| 317 | /* initialize timestamp */ |
| 318 | __touch_watchdog(); |
| 319 | |
| 320 | /* kick off the timer for the hardlockup detector */ |
| 321 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
| 322 | hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()), |
| 323 | HRTIMER_MODE_REL_PINNED); |
| 324 | |
| 325 | set_current_state(TASK_INTERRUPTIBLE); |
| 326 | /* |
| 327 | * Run briefly once per second to reset the softlockup timestamp. |
| 328 | * If this gets delayed for more than 60 seconds then the |
Don Zickus | 26e09c6 | 2010-05-17 18:06:04 -0400 | [diff] [blame] | 329 | * debug-printout triggers in watchdog_timer_fn(). |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 330 | */ |
| 331 | while (!kthread_should_stop()) { |
| 332 | __touch_watchdog(); |
| 333 | schedule(); |
| 334 | |
| 335 | if (kthread_should_stop()) |
| 336 | break; |
| 337 | |
| 338 | set_current_state(TASK_INTERRUPTIBLE); |
| 339 | } |
| 340 | __set_current_state(TASK_RUNNING); |
| 341 | |
| 342 | return 0; |
| 343 | } |
| 344 | |
| 345 | |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 346 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 347 | static int watchdog_nmi_enable(int cpu) |
| 348 | { |
| 349 | struct perf_event_attr *wd_attr; |
| 350 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
| 351 | |
| 352 | /* is it already setup and enabled? */ |
| 353 | if (event && event->state > PERF_EVENT_STATE_OFF) |
| 354 | goto out; |
| 355 | |
| 356 | /* it is setup but not enabled */ |
| 357 | if (event != NULL) |
| 358 | goto out_enable; |
| 359 | |
| 360 | /* Try to register using hardware perf events */ |
| 361 | wd_attr = &wd_hw_attr; |
| 362 | wd_attr->sample_period = hw_nmi_get_sample_period(); |
Matt Helsley | 38a81da | 2010-09-13 13:01:20 -0700 | [diff] [blame] | 363 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 364 | if (!IS_ERR(event)) { |
| 365 | printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n"); |
| 366 | goto out_save; |
| 367 | } |
| 368 | |
Ben Hutchings | 5514237 | 2011-01-02 23:02:42 +0000 | [diff] [blame] | 369 | printk(KERN_ERR "NMI watchdog disabled for cpu%i: unable to create perf event: %ld\n", |
| 370 | cpu, PTR_ERR(event)); |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 371 | return PTR_ERR(event); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 372 | |
| 373 | /* success path */ |
| 374 | out_save: |
| 375 | per_cpu(watchdog_ev, cpu) = event; |
| 376 | out_enable: |
| 377 | perf_event_enable(per_cpu(watchdog_ev, cpu)); |
| 378 | out: |
| 379 | return 0; |
| 380 | } |
| 381 | |
| 382 | static void watchdog_nmi_disable(int cpu) |
| 383 | { |
| 384 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
| 385 | |
| 386 | if (event) { |
| 387 | perf_event_disable(event); |
| 388 | per_cpu(watchdog_ev, cpu) = NULL; |
| 389 | |
| 390 | /* should be in cleanup, but blocks oprofile */ |
| 391 | perf_event_release_kernel(event); |
| 392 | } |
| 393 | return; |
| 394 | } |
| 395 | #else |
| 396 | static int watchdog_nmi_enable(int cpu) { return 0; } |
| 397 | static void watchdog_nmi_disable(int cpu) { return; } |
Frederic Weisbecker | 23637d4 | 2010-05-15 23:15:20 +0200 | [diff] [blame] | 398 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 399 | |
| 400 | /* prepare/enable/disable routines */ |
| 401 | static int watchdog_prepare_cpu(int cpu) |
| 402 | { |
| 403 | struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu); |
| 404 | |
| 405 | WARN_ON(per_cpu(softlockup_watchdog, cpu)); |
| 406 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 407 | hrtimer->function = watchdog_timer_fn; |
| 408 | |
| 409 | return 0; |
| 410 | } |
| 411 | |
| 412 | static int watchdog_enable(int cpu) |
| 413 | { |
| 414 | struct task_struct *p = per_cpu(softlockup_watchdog, cpu); |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 415 | int err; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 416 | |
| 417 | /* enable the perf event */ |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 418 | err = watchdog_nmi_enable(cpu); |
| 419 | if (err) |
| 420 | return err; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 421 | |
| 422 | /* create the watchdog thread */ |
| 423 | if (!p) { |
| 424 | p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu); |
| 425 | if (IS_ERR(p)) { |
| 426 | printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu); |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 427 | return PTR_ERR(p); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 428 | } |
| 429 | kthread_bind(p, cpu); |
| 430 | per_cpu(watchdog_touch_ts, cpu) = 0; |
| 431 | per_cpu(softlockup_watchdog, cpu) = p; |
| 432 | wake_up_process(p); |
| 433 | } |
| 434 | |
Don Zickus | 68d3f1d | 2010-08-31 23:00:07 -0400 | [diff] [blame] | 435 | /* if any cpu succeeds, watchdog is considered enabled for the system */ |
| 436 | watchdog_enabled = 1; |
| 437 | |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 438 | return 0; |
| 439 | } |
| 440 | |
| 441 | static void watchdog_disable(int cpu) |
| 442 | { |
| 443 | struct task_struct *p = per_cpu(softlockup_watchdog, cpu); |
| 444 | struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu); |
| 445 | |
| 446 | /* |
| 447 | * cancel the timer first to stop incrementing the stats |
| 448 | * and waking up the kthread |
| 449 | */ |
| 450 | hrtimer_cancel(hrtimer); |
| 451 | |
| 452 | /* disable the perf event */ |
| 453 | watchdog_nmi_disable(cpu); |
| 454 | |
| 455 | /* stop the watchdog thread */ |
| 456 | if (p) { |
| 457 | per_cpu(softlockup_watchdog, cpu) = NULL; |
| 458 | kthread_stop(p); |
| 459 | } |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 460 | } |
| 461 | |
| 462 | static void watchdog_enable_all_cpus(void) |
| 463 | { |
| 464 | int cpu; |
Kulikov Vasiliy | eb703f9 | 2010-07-05 12:00:54 +0400 | [diff] [blame] | 465 | int result = 0; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 466 | |
| 467 | for_each_online_cpu(cpu) |
| 468 | result += watchdog_enable(cpu); |
| 469 | |
| 470 | if (result) |
| 471 | printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n"); |
| 472 | |
| 473 | } |
| 474 | |
| 475 | static void watchdog_disable_all_cpus(void) |
| 476 | { |
| 477 | int cpu; |
| 478 | |
Stephane Eranian | d9ca07a | 2010-09-14 15:34:01 +0200 | [diff] [blame] | 479 | if (no_watchdog) |
| 480 | return; |
| 481 | |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 482 | for_each_online_cpu(cpu) |
| 483 | watchdog_disable(cpu); |
| 484 | |
| 485 | /* if all watchdogs are disabled, then they are disabled for the system */ |
| 486 | watchdog_enabled = 0; |
| 487 | } |
| 488 | |
| 489 | |
| 490 | /* sysctl functions */ |
| 491 | #ifdef CONFIG_SYSCTL |
| 492 | /* |
| 493 | * proc handler for /proc/sys/kernel/nmi_watchdog |
| 494 | */ |
| 495 | |
| 496 | int proc_dowatchdog_enabled(struct ctl_table *table, int write, |
| 497 | void __user *buffer, size_t *length, loff_t *ppos) |
| 498 | { |
| 499 | proc_dointvec(table, write, buffer, length, ppos); |
| 500 | |
| 501 | if (watchdog_enabled) |
| 502 | watchdog_enable_all_cpus(); |
| 503 | else |
| 504 | watchdog_disable_all_cpus(); |
| 505 | return 0; |
| 506 | } |
| 507 | |
| 508 | int proc_dowatchdog_thresh(struct ctl_table *table, int write, |
| 509 | void __user *buffer, |
| 510 | size_t *lenp, loff_t *ppos) |
| 511 | { |
| 512 | return proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 513 | } |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 514 | #endif /* CONFIG_SYSCTL */ |
| 515 | |
| 516 | |
| 517 | /* |
| 518 | * Create/destroy watchdog threads as CPUs come and go: |
| 519 | */ |
| 520 | static int __cpuinit |
| 521 | cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu) |
| 522 | { |
| 523 | int hotcpu = (unsigned long)hcpu; |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 524 | int err = 0; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 525 | |
| 526 | switch (action) { |
| 527 | case CPU_UP_PREPARE: |
| 528 | case CPU_UP_PREPARE_FROZEN: |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 529 | err = watchdog_prepare_cpu(hotcpu); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 530 | break; |
| 531 | case CPU_ONLINE: |
| 532 | case CPU_ONLINE_FROZEN: |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 533 | err = watchdog_enable(hotcpu); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 534 | break; |
| 535 | #ifdef CONFIG_HOTPLUG_CPU |
| 536 | case CPU_UP_CANCELED: |
| 537 | case CPU_UP_CANCELED_FROZEN: |
| 538 | watchdog_disable(hotcpu); |
| 539 | break; |
| 540 | case CPU_DEAD: |
| 541 | case CPU_DEAD_FROZEN: |
| 542 | watchdog_disable(hotcpu); |
| 543 | break; |
| 544 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 545 | } |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 546 | return notifier_from_errno(err); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 547 | } |
| 548 | |
| 549 | static struct notifier_block __cpuinitdata cpu_nfb = { |
| 550 | .notifier_call = cpu_callback |
| 551 | }; |
| 552 | |
Peter Zijlstra | 004417a | 2010-11-25 18:38:29 +0100 | [diff] [blame] | 553 | void __init lockup_detector_init(void) |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 554 | { |
| 555 | void *cpu = (void *)(long)smp_processor_id(); |
| 556 | int err; |
| 557 | |
| 558 | if (no_watchdog) |
Peter Zijlstra | 004417a | 2010-11-25 18:38:29 +0100 | [diff] [blame] | 559 | return; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 560 | |
| 561 | err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu); |
Akinobu Mita | eac2433 | 2010-08-31 23:00:08 -0400 | [diff] [blame] | 562 | WARN_ON(notifier_to_errno(err)); |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 563 | |
| 564 | cpu_callback(&cpu_nfb, CPU_ONLINE, cpu); |
| 565 | register_cpu_notifier(&cpu_nfb); |
| 566 | |
Peter Zijlstra | 004417a | 2010-11-25 18:38:29 +0100 | [diff] [blame] | 567 | return; |
Don Zickus | 58687ac | 2010-05-07 17:11:44 -0400 | [diff] [blame] | 568 | } |