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 watchdog_thresh = 10;
  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                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  53
  54 static int __init hardlockup_panic_setup(char *str)
  55 {
  56         if (!strncmp(str, "panic", 5))
  57                 hardlockup_panic = 1;
  58         else if (!strncmp(str, "nopanic", 7))
  59                 hardlockup_panic = 0;
  60         else if (!strncmp(str, "0", 1))
  61                 watchdog_enabled = 0;
  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         watchdog_enabled = 0;
  81         return 1;
  82 }
  83 __setup("nowatchdog", nowatchdog_setup);
  84
  85 /* deprecated */
  86 static int __init nosoftlockup_setup(char *str)
  87 {
  88         watchdog_enabled = 0;
  89         return 1;
  90 }
  91 __setup("nosoftlockup", nosoftlockup_setup);
  92 /*  */
  93
  94 /*
  95  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
  96  * lockups can have false positives under extreme conditions. So we generally
  97  * want a higher threshold for soft lockups than for hard lockups. So we couple
  98  * the thresholds with a factor: we make the soft threshold twice the amount of
  99  * time the hard threshold is.
 100  */
 101 static int get_softlockup_thresh(void)
 102 {
 103         return watchdog_thresh * 2;
 104 }
 105
 106 /*
 107  * Returns seconds, approximately.  We don't need nanosecond
 108  * resolution, and we don't need to waste time with a big divide when
 109  * 2^30ns == 1.074s.
 110  */
 111 static unsigned long get_timestamp(int this_cpu)
 112 {
 113         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 114 }
 115
 116 static u64 get_sample_period(void)
 117 {
 118         /*
 119          * convert watchdog_thresh from seconds to ns
 120          * the divide by 5 is to give hrtimer 5 chances to
 121          * increment before the hardlockup detector generates
 122          * a warning
 123          */
 124         return get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 125 }
 126
 127 /* Commands for resetting the watchdog */
 128 static void __touch_watchdog(void)
 129 {
 130         int this_cpu = smp_processor_id();
 131
 132         __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
 133 }
 134
 135 void touch_softlockup_watchdog(void)
 136 {
 137         __this_cpu_write(watchdog_touch_ts, 0);
 138 }
 139 EXPORT_SYMBOL(touch_softlockup_watchdog);
 140
 141 void touch_all_softlockup_watchdogs(void)
 142 {
 143         int cpu;
 144
 145         /*
 146          * this is done lockless
 147          * do we care if a 0 races with a timestamp?
 148          * all it means is the softlock check starts one cycle later
 149          */
 150         for_each_online_cpu(cpu)
 151                 per_cpu(watchdog_touch_ts, cpu) = 0;
 152 }
 153
 154 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 155 void touch_nmi_watchdog(void)
 156 {
 157         if (watchdog_enabled) {
 158                 unsigned cpu;
 159
 160                 for_each_present_cpu(cpu) {
 161                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
 162                                 per_cpu(watchdog_nmi_touch, cpu) = true;
 163                 }
 164         }
 165         touch_softlockup_watchdog();
 166 }
 167 EXPORT_SYMBOL(touch_nmi_watchdog);
 168
 169 #endif
 170
 171 void touch_softlockup_watchdog_sync(void)
 172 {
 173         __raw_get_cpu_var(softlockup_touch_sync) = true;
 174         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 175 }
 176
 177 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 178 /* watchdog detector functions */
 179 static int is_hardlockup(void)
 180 {
 181         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 182
 183         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 184                 return 1;
 185
 186         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 187         return 0;
 188 }
 189 #endif
 190
 191 static int is_softlockup(unsigned long touch_ts)
 192 {
 193         unsigned long now = get_timestamp(smp_processor_id());
 194
 195         /* Warn about unreasonable delays: */
 196         if (time_after(now, touch_ts + get_softlockup_thresh()))
 197                 return now - touch_ts;
 198
 199         return 0;
 200 }
 201
 202 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 203
 204 static struct perf_event_attr wd_hw_attr = {
 205         .type           = PERF_TYPE_HARDWARE,
 206         .config         = PERF_COUNT_HW_CPU_CYCLES,
 207         .size           = sizeof(struct perf_event_attr),
 208         .pinned         = 1,
 209         .disabled       = 1,
 210 };
 211
 212 /* Callback function for perf event subsystem */
 213 static void watchdog_overflow_callback(struct perf_event *event,
 214                  struct perf_sample_data *data,
 215                  struct pt_regs *regs)
 216 {
 217         /* Ensure the watchdog never gets throttled */
 218         event->hw.interrupts = 0;
 219
 220         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 221                 __this_cpu_write(watchdog_nmi_touch, false);
 222                 return;
 223         }
 224
 225         /* check for a hardlockup
 226          * This is done by making sure our timer interrupt
 227          * is incrementing.  The timer interrupt should have
 228          * fired multiple times before we overflow'd.  If it hasn't
 229          * then this is a good indication the cpu is stuck
 230          */
 231         if (is_hardlockup()) {
 232                 int this_cpu = smp_processor_id();
 233
 234                 /* only print hardlockups once */
 235                 if (__this_cpu_read(hard_watchdog_warn) == true)
 236                         return;
 237
 238                 if (hardlockup_panic)
 239                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 240                 else
 241                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 242
 243                 __this_cpu_write(hard_watchdog_warn, true);
 244                 return;
 245         }
 246
 247         __this_cpu_write(hard_watchdog_warn, false);
 248         return;
 249 }
 250 static void watchdog_interrupt_count(void)
 251 {
 252         __this_cpu_inc(hrtimer_interrupts);
 253 }
 254 #else
 255 static inline void watchdog_interrupt_count(void) { return; }
 256 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 257
 258 /* watchdog kicker functions */
 259 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 260 {
 261         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 262         struct pt_regs *regs = get_irq_regs();
 263         int duration;
 264
 265         /* kick the hardlockup detector */
 266         watchdog_interrupt_count();
 267
 268         /* kick the softlockup detector */
 269         wake_up_process(__this_cpu_read(softlockup_watchdog));
 270
 271         /* .. and repeat */
 272         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
 273
 274         if (touch_ts == 0) {
 275                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 276                         /*
 277                          * If the time stamp was touched atomically
 278                          * make sure the scheduler tick is up to date.
 279                          */
 280                         __this_cpu_write(softlockup_touch_sync, false);
 281                         sched_clock_tick();
 282                 }
 283                 __touch_watchdog();
 284                 return HRTIMER_RESTART;
 285         }
 286
 287         /* check for a softlockup
 288          * This is done by making sure a high priority task is
 289          * being scheduled.  The task touches the watchdog to
 290          * indicate it is getting cpu time.  If it hasn't then
 291          * this is a good indication some task is hogging the cpu
 292          */
 293         duration = is_softlockup(touch_ts);
 294         if (unlikely(duration)) {
 295                 /* only warn once */
 296                 if (__this_cpu_read(soft_watchdog_warn) == true)
 297                         return HRTIMER_RESTART;
 298
 299                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 300                         smp_processor_id(), duration,
 301                         current->comm, task_pid_nr(current));
 302                 print_modules();
 303                 print_irqtrace_events(current);
 304                 if (regs)
 305                         show_regs(regs);
 306                 else
 307                         dump_stack();
 308
 309                 if (softlockup_panic)
 310                         panic("softlockup: hung tasks");
 311                 __this_cpu_write(soft_watchdog_warn, true);
 312         } else
 313                 __this_cpu_write(soft_watchdog_warn, false);
 314
 315         return HRTIMER_RESTART;
 316 }
 317
 318
 319 /*
 320  * The watchdog thread - touches the timestamp.
 321  */
 322 static int watchdog(void *unused)
 323 {
 324         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 325         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 326
 327         sched_setscheduler(current, SCHED_FIFO, &param);
 328
 329         /* initialize timestamp */
 330         __touch_watchdog();
 331
 332         /* kick off the timer for the hardlockup detector */
 333         /* done here because hrtimer_start can only pin to smp_processor_id() */
 334         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
 335                       HRTIMER_MODE_REL_PINNED);
 336
 337         set_current_state(TASK_INTERRUPTIBLE);
 338         /*
 339          * Run briefly once per second to reset the softlockup timestamp.
 340          * If this gets delayed for more than 60 seconds then the
 341          * debug-printout triggers in watchdog_timer_fn().
 342          */
 343         while (!kthread_should_stop()) {
 344                 __touch_watchdog();
 345                 schedule();
 346
 347                 if (kthread_should_stop())
 348                         break;
 349
 350                 set_current_state(TASK_INTERRUPTIBLE);
 351         }
 352         __set_current_state(TASK_RUNNING);
 353         param.sched_priority = 0;
 354         sched_setscheduler(current, SCHED_NORMAL, &param);
 355         return 0;
 356 }
 357
 358
 359 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 360 static int watchdog_nmi_enable(int cpu)
 361 {
 362         struct perf_event_attr *wd_attr;
 363         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 364
 365         /* is it already setup and enabled? */
 366         if (event && event->state > PERF_EVENT_STATE_OFF)
 367                 goto out;
 368
 369         /* it is setup but not enabled */
 370         if (event != NULL)
 371                 goto out_enable;
 372
 373         wd_attr = &wd_hw_attr;
 374         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 375
 376         /* Try to register using hardware perf events */
 377         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 378         if (!IS_ERR(event)) {
 379                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
 380                 goto out_save;
 381         }
 382
 383
 384         /* vary the KERN level based on the returned errno */
 385         if (PTR_ERR(event) == -EOPNOTSUPP)
 386                 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 387         else if (PTR_ERR(event) == -ENOENT)
 388                 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
 389         else
 390                 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
 391         return PTR_ERR(event);
 392
 393         /* success path */
 394 out_save:
 395         per_cpu(watchdog_ev, cpu) = event;
 396 out_enable:
 397         perf_event_enable(per_cpu(watchdog_ev, cpu));
 398 out:
 399         return 0;
 400 }
 401
 402 static void watchdog_nmi_disable(int cpu)
 403 {
 404         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 405
 406         if (event) {
 407                 perf_event_disable(event);
 408                 per_cpu(watchdog_ev, cpu) = NULL;
 409
 410                 /* should be in cleanup, but blocks oprofile */
 411                 perf_event_release_kernel(event);
 412         }
 413         return;
 414 }
 415 #else
 416 static int watchdog_nmi_enable(int cpu) { return 0; }
 417 static void watchdog_nmi_disable(int cpu) { return; }
 418 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 419
 420 /* prepare/enable/disable routines */
 421 static void watchdog_prepare_cpu(int cpu)
 422 {
 423         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 424
 425         WARN_ON(per_cpu(softlockup_watchdog, cpu));
 426         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 427         hrtimer->function = watchdog_timer_fn;
 428 }
 429
 430 static int watchdog_enable(int cpu)
 431 {
 432         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 433         int err = 0;
 434
 435         /* enable the perf event */
 436         err = watchdog_nmi_enable(cpu);
 437
 438         /* Regardless of err above, fall through and start softlockup */
 439
 440         /* create the watchdog thread */
 441         if (!p) {
 442                 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
 443                 if (IS_ERR(p)) {
 444                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
 445                         if (!err) {
 446                                 /* if hardlockup hasn't already set this */
 447                                 err = PTR_ERR(p);
 448                                 /* and disable the perf event */
 449                                 watchdog_nmi_disable(cpu);
 450                         }
 451                         goto out;
 452                 }
 453                 kthread_bind(p, cpu);
 454                 per_cpu(watchdog_touch_ts, cpu) = 0;
 455                 per_cpu(softlockup_watchdog, cpu) = p;
 456                 wake_up_process(p);
 457         }
 458
 459 out:
 460         return err;
 461 }
 462
 463 static void watchdog_disable(int cpu)
 464 {
 465         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 466         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 467
 468         /*
 469          * cancel the timer first to stop incrementing the stats
 470          * and waking up the kthread
 471          */
 472         hrtimer_cancel(hrtimer);
 473
 474         /* disable the perf event */
 475         watchdog_nmi_disable(cpu);
 476
 477         /* stop the watchdog thread */
 478         if (p) {
 479                 per_cpu(softlockup_watchdog, cpu) = NULL;
 480                 kthread_stop(p);
 481         }
 482 }
 483
 484 /* sysctl functions */
 485 #ifdef CONFIG_SYSCTL
 486 static void watchdog_enable_all_cpus(void)
 487 {
 488         int cpu;
 489
 490         watchdog_enabled = 0;
 491
 492         for_each_online_cpu(cpu)
 493                 if (!watchdog_enable(cpu))
 494                         /* if any cpu succeeds, watchdog is considered
 495                            enabled for the system */
 496                         watchdog_enabled = 1;
 497
 498         if (!watchdog_enabled)
 499                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
 500
 501 }
 502
 503 static void watchdog_disable_all_cpus(void)
 504 {
 505         int cpu;
 506
 507         for_each_online_cpu(cpu)
 508                 watchdog_disable(cpu);
 509
 510         /* if all watchdogs are disabled, then they are disabled for the system */
 511         watchdog_enabled = 0;
 512 }
 513
 514
 515 /*
 516  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 517  */
 518
 519 int proc_dowatchdog(struct ctl_table *table, int write,
 520                     void __user *buffer, size_t *lenp, loff_t *ppos)
 521 {
 522         int ret;
 523
 524         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 525         if (ret || !write)
 526                 goto out;
 527
 528         if (watchdog_enabled && watchdog_thresh)
 529                 watchdog_enable_all_cpus();
 530         else
 531                 watchdog_disable_all_cpus();
 532
 533 out:
 534         return ret;
 535 }
 536 #endif /* CONFIG_SYSCTL */
 537
 538
 539 /*
 540  * Create/destroy watchdog threads as CPUs come and go:
 541  */
 542 static int __cpuinit
 543 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 544 {
 545         int hotcpu = (unsigned long)hcpu;
 546
 547         switch (action) {
 548         case CPU_UP_PREPARE:
 549         case CPU_UP_PREPARE_FROZEN:
 550                 watchdog_prepare_cpu(hotcpu);
 551                 break;
 552         case CPU_ONLINE:
 553         case CPU_ONLINE_FROZEN:
 554                 if (watchdog_enabled)
 555                         watchdog_enable(hotcpu);
 556                 break;
 557 #ifdef CONFIG_HOTPLUG_CPU
 558         case CPU_UP_CANCELED:
 559         case CPU_UP_CANCELED_FROZEN:
 560                 watchdog_disable(hotcpu);
 561                 break;
 562         case CPU_DEAD:
 563         case CPU_DEAD_FROZEN:
 564                 watchdog_disable(hotcpu);
 565                 break;
 566 #endif /* CONFIG_HOTPLUG_CPU */
 567         }
 568
 569         /*
 570          * hardlockup and softlockup are not important enough
 571          * to block cpu bring up.  Just always succeed and
 572          * rely on printk output to flag problems.
 573          */
 574         return NOTIFY_OK;
 575 }
 576
 577 static struct notifier_block __cpuinitdata cpu_nfb = {
 578         .notifier_call = cpu_callback
 579 };
 580
 581 void __init lockup_detector_init(void)
 582 {
 583         void *cpu = (void *)(long)smp_processor_id();
 584         int err;
 585
 586         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 587         WARN_ON(notifier_to_errno(err));
 588
 589         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 590         register_cpu_notifier(&cpu_nfb);
 591
 592         return;
 593 }