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 = 1;
  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 /* boot commands */
  47 /*
  48  * Should we panic when a soft-lockup or hard-lockup occurs:
  49  */
  50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  51 static int hardlockup_panic;
  52
  53 static int __init hardlockup_panic_setup(char *str)
  54 {
  55         if (!strncmp(str, "panic", 5))
  56                 hardlockup_panic = 1;
  57         else if (!strncmp(str, "0", 1))
  58                 watchdog_enabled = 0;
  59         return 1;
  60 }
  61 __setup("nmi_watchdog=", hardlockup_panic_setup);
  62 #endif
  63
  64 unsigned int __read_mostly softlockup_panic =
  65                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  66
  67 static int __init softlockup_panic_setup(char *str)
  68 {
  69         softlockup_panic = simple_strtoul(str, NULL, 0);
  70
  71         return 1;
  72 }
  73 __setup("softlockup_panic=", softlockup_panic_setup);
  74
  75 static int __init nowatchdog_setup(char *str)
  76 {
  77         watchdog_enabled = 0;
  78         return 1;
  79 }
  80 __setup("nowatchdog", nowatchdog_setup);
  81
  82 /* deprecated */
  83 static int __init nosoftlockup_setup(char *str)
  84 {
  85         watchdog_enabled = 0;
  86         return 1;
  87 }
  88 __setup("nosoftlockup", nosoftlockup_setup);
  89 /*  */
  90
  91
  92 /*
  93  * Returns seconds, approximately.  We don't need nanosecond
  94  * resolution, and we don't need to waste time with a big divide when
  95  * 2^30ns == 1.074s.
  96  */
  97 static unsigned long get_timestamp(int this_cpu)
  98 {
  99         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 100 }
 101
 102 static unsigned long get_sample_period(void)
 103 {
 104         /*
 105          * convert softlockup_thresh from seconds to ns
 106          * the divide by 5 is to give hrtimer 5 chances to
 107          * increment before the hardlockup detector generates
 108          * a warning
 109          */
 110         return softlockup_thresh / 5 * NSEC_PER_SEC;
 111 }
 112
 113 /* Commands for resetting the watchdog */
 114 static void __touch_watchdog(void)
 115 {
 116         int this_cpu = smp_processor_id();
 117
 118         __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
 119 }
 120
 121 void touch_softlockup_watchdog(void)
 122 {
 123         __this_cpu_write(watchdog_touch_ts, 0);
 124 }
 125 EXPORT_SYMBOL(touch_softlockup_watchdog);
 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 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 141 void touch_nmi_watchdog(void)
 142 {
 143         if (watchdog_enabled) {
 144                 unsigned cpu;
 145
 146                 for_each_present_cpu(cpu) {
 147                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
 148                                 per_cpu(watchdog_nmi_touch, cpu) = true;
 149                 }
 150         }
 151         touch_softlockup_watchdog();
 152 }
 153 EXPORT_SYMBOL(touch_nmi_watchdog);
 154
 155 #endif
 156
 157 void touch_softlockup_watchdog_sync(void)
 158 {
 159         __raw_get_cpu_var(softlockup_touch_sync) = true;
 160         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 161 }
 162
 163 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 164 /* watchdog detector functions */
 165 static int is_hardlockup(void)
 166 {
 167         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 168
 169         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 170                 return 1;
 171
 172         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 173         return 0;
 174 }
 175 #endif
 176
 177 static int is_softlockup(unsigned long touch_ts)
 178 {
 179         unsigned long now = get_timestamp(smp_processor_id());
 180
 181         /* Warn about unreasonable delays: */
 182         if (time_after(now, touch_ts + softlockup_thresh))
 183                 return now - touch_ts;
 184
 185         return 0;
 186 }
 187
 188 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 189 static struct perf_event_attr wd_hw_attr = {
 190         .type           = PERF_TYPE_HARDWARE,
 191         .config         = PERF_COUNT_HW_CPU_CYCLES,
 192         .size           = sizeof(struct perf_event_attr),
 193         .pinned         = 1,
 194         .disabled       = 1,
 195 };
 196
 197 /* Callback function for perf event subsystem */
 198 static void watchdog_overflow_callback(struct perf_event *event, int nmi,
 199                  struct perf_sample_data *data,
 200                  struct pt_regs *regs)
 201 {
 202         /* Ensure the watchdog never gets throttled */
 203         event->hw.interrupts = 0;
 204
 205         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 206                 __this_cpu_write(watchdog_nmi_touch, false);
 207                 return;
 208         }
 209
 210         /* check for a hardlockup
 211          * This is done by making sure our timer interrupt
 212          * is incrementing.  The timer interrupt should have
 213          * fired multiple times before we overflow'd.  If it hasn't
 214          * then this is a good indication the cpu is stuck
 215          */
 216         if (is_hardlockup()) {
 217                 int this_cpu = smp_processor_id();
 218
 219                 /* only print hardlockups once */
 220                 if (__this_cpu_read(hard_watchdog_warn) == true)
 221                         return;
 222
 223                 if (hardlockup_panic)
 224                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 225                 else
 226                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 227
 228                 __this_cpu_write(hard_watchdog_warn, true);
 229                 return;
 230         }
 231
 232         __this_cpu_write(hard_watchdog_warn, false);
 233         return;
 234 }
 235 static void watchdog_interrupt_count(void)
 236 {
 237         __this_cpu_inc(hrtimer_interrupts);
 238 }
 239 #else
 240 static inline void watchdog_interrupt_count(void) { return; }
 241 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 242
 243 /* watchdog kicker functions */
 244 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 245 {
 246         unsigned long touch_ts = __this_cpu_read(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(__this_cpu_read(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(__this_cpu_read(softlockup_touch_sync))) {
 261                         /*
 262                          * If the time stamp was touched atomically
 263                          * make sure the scheduler tick is up to date.
 264                          */
 265                         __this_cpu_write(softlockup_touch_sync, 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);
 279         if (unlikely(duration)) {
 280                 /* only warn once */
 281                 if (__this_cpu_read(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                         smp_processor_id(), 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                 __this_cpu_write(soft_watchdog_warn, true);
 297         } else
 298                 __this_cpu_write(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 *unused)
 308 {
 309         static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 310         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 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 watchdog_timer_fn().
 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_HARDLOCKUP_DETECTOR
 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, NULL, 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
 367         /* vary the KERN level based on the returned errno */
 368         if (PTR_ERR(event) == -EOPNOTSUPP)
 369                 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 370         else if (PTR_ERR(event) == -ENOENT)
 371                 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
 372         else
 373                 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
 374         return PTR_ERR(event);
 375
 376         /* success path */
 377 out_save:
 378         per_cpu(watchdog_ev, cpu) = event;
 379 out_enable:
 380         perf_event_enable(per_cpu(watchdog_ev, cpu));
 381 out:
 382         return 0;
 383 }
 384
 385 static void watchdog_nmi_disable(int cpu)
 386 {
 387         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 388
 389         if (event) {
 390                 perf_event_disable(event);
 391                 per_cpu(watchdog_ev, cpu) = NULL;
 392
 393                 /* should be in cleanup, but blocks oprofile */
 394                 perf_event_release_kernel(event);
 395         }
 396         return;
 397 }
 398 #else
 399 static int watchdog_nmi_enable(int cpu) { return 0; }
 400 static void watchdog_nmi_disable(int cpu) { return; }
 401 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 402
 403 /* prepare/enable/disable routines */
 404 static int watchdog_prepare_cpu(int cpu)
 405 {
 406         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 407
 408         WARN_ON(per_cpu(softlockup_watchdog, cpu));
 409         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 410         hrtimer->function = watchdog_timer_fn;
 411
 412         return 0;
 413 }
 414
 415 static int watchdog_enable(int cpu)
 416 {
 417         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 418         int err;
 419
 420         /* enable the perf event */
 421         err = watchdog_nmi_enable(cpu);
 422         if (err)
 423                 return err;
 424
 425         /* create the watchdog thread */
 426         if (!p) {
 427                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
 428                 if (IS_ERR(p)) {
 429                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
 430                         return PTR_ERR(p);
 431                 }
 432                 kthread_bind(p, cpu);
 433                 per_cpu(watchdog_touch_ts, cpu) = 0;
 434                 per_cpu(softlockup_watchdog, cpu) = p;
 435                 wake_up_process(p);
 436         }
 437
 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         }
 460 }
 461
 462 static void watchdog_enable_all_cpus(void)
 463 {
 464         int cpu;
 465
 466         watchdog_enabled = 0;
 467
 468         for_each_online_cpu(cpu)
 469                 if (!watchdog_enable(cpu))
 470                         /* if any cpu succeeds, watchdog is considered
 471                            enabled for the system */
 472                         watchdog_enabled = 1;
 473
 474         if (!watchdog_enabled)
 475                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
 476
 477 }
 478
 479 static void watchdog_disable_all_cpus(void)
 480 {
 481         int cpu;
 482
 483         for_each_online_cpu(cpu)
 484                 watchdog_disable(cpu);
 485
 486         /* if all watchdogs are disabled, then they are disabled for the system */
 487         watchdog_enabled = 0;
 488 }
 489
 490
 491 /* sysctl functions */
 492 #ifdef CONFIG_SYSCTL
 493 /*
 494  * proc handler for /proc/sys/kernel/nmi_watchdog
 495  */
 496
 497 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
 498                      void __user *buffer, size_t *length, loff_t *ppos)
 499 {
 500         proc_dointvec(table, write, buffer, length, ppos);
 501
 502         if (write) {
 503                 if (watchdog_enabled)
 504                         watchdog_enable_all_cpus();
 505                 else
 506                         watchdog_disable_all_cpus();
 507         }
 508         return 0;
 509 }
 510
 511 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
 512                              void __user *buffer,
 513                              size_t *lenp, loff_t *ppos)
 514 {
 515         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 516 }
 517 #endif /* CONFIG_SYSCTL */
 518
 519
 520 /*
 521  * Create/destroy watchdog threads as CPUs come and go:
 522  */
 523 static int __cpuinit
 524 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 525 {
 526         int hotcpu = (unsigned long)hcpu;
 527         int err = 0;
 528
 529         switch (action) {
 530         case CPU_UP_PREPARE:
 531         case CPU_UP_PREPARE_FROZEN:
 532                 err = watchdog_prepare_cpu(hotcpu);
 533                 break;
 534         case CPU_ONLINE:
 535         case CPU_ONLINE_FROZEN:
 536                 if (watchdog_enabled)
 537                         err = watchdog_enable(hotcpu);
 538                 break;
 539 #ifdef CONFIG_HOTPLUG_CPU
 540         case CPU_UP_CANCELED:
 541         case CPU_UP_CANCELED_FROZEN:
 542                 watchdog_disable(hotcpu);
 543                 break;
 544         case CPU_DEAD:
 545         case CPU_DEAD_FROZEN:
 546                 watchdog_disable(hotcpu);
 547                 break;
 548 #endif /* CONFIG_HOTPLUG_CPU */
 549         }
 550         return notifier_from_errno(err);
 551 }
 552
 553 static struct notifier_block __cpuinitdata cpu_nfb = {
 554         .notifier_call = cpu_callback
 555 };
 556
 557 void __init lockup_detector_init(void)
 558 {
 559         void *cpu = (void *)(long)smp_processor_id();
 560         int err;
 561
 562         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 563         WARN_ON(notifier_to_errno(err));
 564
 565         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 566         register_cpu_notifier(&cpu_nfb);
 567
 568         return;
 569 }