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