kernel/watchdog.c

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