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  * Note: Most of this code is borrowed heavily from the original softlockup
   7  * detector, so thanks to Ingo for the initial implementation.
   8  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
   9  * to those contributors as well.
  10  */
  11
  12 #define pr_fmt(fmt) "NMI watchdog: " fmt
  13
  14 #include <linux/mm.h>
  15 #include <linux/cpu.h>
  16 #include <linux/nmi.h>
  17 #include <linux/init.h>
  18 #include <linux/delay.h>
  19 #include <linux/freezer.h>
  20 #include <linux/kthread.h>
  21 #include <linux/lockdep.h>
  22 #include <linux/notifier.h>
  23 #include <linux/module.h>
  24 #include <linux/sysctl.h>
  25 #include <linux/smpboot.h>
  26 #include <linux/sched/rt.h>
  27
  28 #include <asm/irq_regs.h>
  29 #include <linux/kvm_para.h>
  30 #include <linux/perf_event.h>
  31
  32 int watchdog_user_enabled = 1;
  33 int __read_mostly watchdog_thresh = 10;
  34 #ifdef CONFIG_SMP
  35 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  36 #else
  37 #define sysctl_softlockup_all_cpu_backtrace 0
  38 #endif
  39
  40 static int __read_mostly watchdog_running;
  41 static u64 __read_mostly sample_period;
  42
  43 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  44 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  45 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  46 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  47 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  48 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  49 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
  50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  51 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  52 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  53 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  54 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  55 #endif
  56 static unsigned long soft_lockup_nmi_warn;
  57
  58 /* boot commands */
  59 /*
  60  * Should we panic when a soft-lockup or hard-lockup occurs:
  61  */
  62 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  63 static int hardlockup_panic =
  64                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  65
  66 static int __init hardlockup_panic_setup(char *str)
  67 {
  68         if (!strncmp(str, "panic", 5))
  69                 hardlockup_panic = 1;
  70         else if (!strncmp(str, "nopanic", 7))
  71                 hardlockup_panic = 0;
  72         else if (!strncmp(str, "0", 1))
  73                 watchdog_user_enabled = 0;
  74         return 1;
  75 }
  76 __setup("nmi_watchdog=", hardlockup_panic_setup);
  77 #endif
  78
  79 unsigned int __read_mostly softlockup_panic =
  80                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  81
  82 static int __init softlockup_panic_setup(char *str)
  83 {
  84         softlockup_panic = simple_strtoul(str, NULL, 0);
  85
  86         return 1;
  87 }
  88 __setup("softlockup_panic=", softlockup_panic_setup);
  89
  90 static int __init nowatchdog_setup(char *str)
  91 {
  92         watchdog_user_enabled = 0;
  93         return 1;
  94 }
  95 __setup("nowatchdog", nowatchdog_setup);
  96
  97 /* deprecated */
  98 static int __init nosoftlockup_setup(char *str)
  99 {
 100         watchdog_user_enabled = 0;
 101         return 1;
 102 }
 103 __setup("nosoftlockup", nosoftlockup_setup);
 104 /*  */
 105 #ifdef CONFIG_SMP
 106 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 107 {
 108         sysctl_softlockup_all_cpu_backtrace =
 109                 !!simple_strtol(str, NULL, 0);
 110         return 1;
 111 }
 112 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 113 #endif
 114
 115 /*
 116  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 117  * lockups can have false positives under extreme conditions. So we generally
 118  * want a higher threshold for soft lockups than for hard lockups. So we couple
 119  * the thresholds with a factor: we make the soft threshold twice the amount of
 120  * time the hard threshold is.
 121  */
 122 static int get_softlockup_thresh(void)
 123 {
 124         return watchdog_thresh * 2;
 125 }
 126
 127 /*
 128  * Returns seconds, approximately.  We don't need nanosecond
 129  * resolution, and we don't need to waste time with a big divide when
 130  * 2^30ns == 1.074s.
 131  */
 132 static unsigned long get_timestamp(void)
 133 {
 134         return local_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 135 }
 136
 137 static void set_sample_period(void)
 138 {
 139         /*
 140          * convert watchdog_thresh from seconds to ns
 141          * the divide by 5 is to give hrtimer several chances (two
 142          * or three with the current relation between the soft
 143          * and hard thresholds) to increment before the
 144          * hardlockup detector generates a warning
 145          */
 146         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 147 }
 148
 149 /* Commands for resetting the watchdog */
 150 static void __touch_watchdog(void)
 151 {
 152         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 153 }
 154
 155 void touch_softlockup_watchdog(void)
 156 {
 157         /*
 158          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 159          * gets zeroed here, so use the raw_ operation.
 160          */
 161         raw_cpu_write(watchdog_touch_ts, 0);
 162 }
 163 EXPORT_SYMBOL(touch_softlockup_watchdog);
 164
 165 void touch_all_softlockup_watchdogs(void)
 166 {
 167         int cpu;
 168
 169         /*
 170          * this is done lockless
 171          * do we care if a 0 races with a timestamp?
 172          * all it means is the softlock check starts one cycle later
 173          */
 174         for_each_online_cpu(cpu)
 175                 per_cpu(watchdog_touch_ts, cpu) = 0;
 176 }
 177
 178 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 179 void touch_nmi_watchdog(void)
 180 {
 181         /*
 182          * Using __raw here because some code paths have
 183          * preemption enabled.  If preemption is enabled
 184          * then interrupts should be enabled too, in which
 185          * case we shouldn't have to worry about the watchdog
 186          * going off.
 187          */
 188         __raw_get_cpu_var(watchdog_nmi_touch) = true;
 189         touch_softlockup_watchdog();
 190 }
 191 EXPORT_SYMBOL(touch_nmi_watchdog);
 192
 193 #endif
 194
 195 void touch_softlockup_watchdog_sync(void)
 196 {
 197         __raw_get_cpu_var(softlockup_touch_sync) = true;
 198         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 199 }
 200
 201 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 202 /* watchdog detector functions */
 203 static int is_hardlockup(void)
 204 {
 205         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 206
 207         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 208                 return 1;
 209
 210         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 211         return 0;
 212 }
 213 #endif
 214
 215 static int is_softlockup(unsigned long touch_ts)
 216 {
 217         unsigned long now = get_timestamp();
 218
 219         /* Warn about unreasonable delays: */
 220         if (time_after(now, touch_ts + get_softlockup_thresh()))
 221                 return now - touch_ts;
 222
 223         return 0;
 224 }
 225
 226 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 227
 228 static struct perf_event_attr wd_hw_attr = {
 229         .type           = PERF_TYPE_HARDWARE,
 230         .config         = PERF_COUNT_HW_CPU_CYCLES,
 231         .size           = sizeof(struct perf_event_attr),
 232         .pinned         = 1,
 233         .disabled       = 1,
 234 };
 235
 236 /* Callback function for perf event subsystem */
 237 static void watchdog_overflow_callback(struct perf_event *event,
 238                  struct perf_sample_data *data,
 239                  struct pt_regs *regs)
 240 {
 241         /* Ensure the watchdog never gets throttled */
 242         event->hw.interrupts = 0;
 243
 244         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 245                 __this_cpu_write(watchdog_nmi_touch, false);
 246                 return;
 247         }
 248
 249         /* check for a hardlockup
 250          * This is done by making sure our timer interrupt
 251          * is incrementing.  The timer interrupt should have
 252          * fired multiple times before we overflow'd.  If it hasn't
 253          * then this is a good indication the cpu is stuck
 254          */
 255         if (is_hardlockup()) {
 256                 int this_cpu = smp_processor_id();
 257
 258                 /* only print hardlockups once */
 259                 if (__this_cpu_read(hard_watchdog_warn) == true)
 260                         return;
 261
 262                 if (hardlockup_panic)
 263                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 264                 else
 265                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 266
 267                 __this_cpu_write(hard_watchdog_warn, true);
 268                 return;
 269         }
 270
 271         __this_cpu_write(hard_watchdog_warn, false);
 272         return;
 273 }
 274 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 275
 276 static void watchdog_interrupt_count(void)
 277 {
 278         __this_cpu_inc(hrtimer_interrupts);
 279 }
 280
 281 static int watchdog_nmi_enable(unsigned int cpu);
 282 static void watchdog_nmi_disable(unsigned int cpu);
 283
 284 /* watchdog kicker functions */
 285 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 286 {
 287         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 288         struct pt_regs *regs = get_irq_regs();
 289         int duration;
 290         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 291
 292         /* kick the hardlockup detector */
 293         watchdog_interrupt_count();
 294
 295         /* kick the softlockup detector */
 296         wake_up_process(__this_cpu_read(softlockup_watchdog));
 297
 298         /* .. and repeat */
 299         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 300
 301         if (touch_ts == 0) {
 302                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 303                         /*
 304                          * If the time stamp was touched atomically
 305                          * make sure the scheduler tick is up to date.
 306                          */
 307                         __this_cpu_write(softlockup_touch_sync, false);
 308                         sched_clock_tick();
 309                 }
 310
 311                 /* Clear the guest paused flag on watchdog reset */
 312                 kvm_check_and_clear_guest_paused();
 313                 __touch_watchdog();
 314                 return HRTIMER_RESTART;
 315         }
 316
 317         /* check for a softlockup
 318          * This is done by making sure a high priority task is
 319          * being scheduled.  The task touches the watchdog to
 320          * indicate it is getting cpu time.  If it hasn't then
 321          * this is a good indication some task is hogging the cpu
 322          */
 323         duration = is_softlockup(touch_ts);
 324         if (unlikely(duration)) {
 325                 /*
 326                  * If a virtual machine is stopped by the host it can look to
 327                  * the watchdog like a soft lockup, check to see if the host
 328                  * stopped the vm before we issue the warning
 329                  */
 330                 if (kvm_check_and_clear_guest_paused())
 331                         return HRTIMER_RESTART;
 332
 333                 /* only warn once */
 334                 if (__this_cpu_read(soft_watchdog_warn) == true)
 335                         return HRTIMER_RESTART;
 336
 337                 if (softlockup_all_cpu_backtrace) {
 338                         /* Prevent multiple soft-lockup reports if one cpu is already
 339                          * engaged in dumping cpu back traces
 340                          */
 341                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 342                                 /* Someone else will report us. Let's give up */
 343                                 __this_cpu_write(soft_watchdog_warn, true);
 344                                 return HRTIMER_RESTART;
 345                         }
 346                 }
 347
 348                 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 349                         smp_processor_id(), duration,
 350                         current->comm, task_pid_nr(current));
 351                 print_modules();
 352                 print_irqtrace_events(current);
 353                 if (regs)
 354                         show_regs(regs);
 355                 else
 356                         dump_stack();
 357
 358                 if (softlockup_all_cpu_backtrace) {
 359                         /* Avoid generating two back traces for current
 360                          * given that one is already made above
 361                          */
 362                         trigger_allbutself_cpu_backtrace();
 363
 364                         clear_bit(0, &soft_lockup_nmi_warn);
 365                         /* Barrier to sync with other cpus */
 366                         smp_mb__after_atomic();
 367                 }
 368
 369                 if (softlockup_panic)
 370                         panic("softlockup: hung tasks");
 371                 __this_cpu_write(soft_watchdog_warn, true);
 372         } else
 373                 __this_cpu_write(soft_watchdog_warn, false);
 374
 375         return HRTIMER_RESTART;
 376 }
 377
 378 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 379 {
 380         struct sched_param param = { .sched_priority = prio };
 381
 382         sched_setscheduler(current, policy, &param);
 383 }
 384
 385 static void watchdog_enable(unsigned int cpu)
 386 {
 387         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 388
 389         /* kick off the timer for the hardlockup detector */
 390         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 391         hrtimer->function = watchdog_timer_fn;
 392
 393         /* Enable the perf event */
 394         watchdog_nmi_enable(cpu);
 395
 396         /* done here because hrtimer_start can only pin to smp_processor_id() */
 397         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 398                       HRTIMER_MODE_REL_PINNED);
 399
 400         /* initialize timestamp */
 401         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 402         __touch_watchdog();
 403 }
 404
 405 static void watchdog_disable(unsigned int cpu)
 406 {
 407         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 408
 409         watchdog_set_prio(SCHED_NORMAL, 0);
 410         hrtimer_cancel(hrtimer);
 411         /* disable the perf event */
 412         watchdog_nmi_disable(cpu);
 413 }
 414
 415 static void watchdog_cleanup(unsigned int cpu, bool online)
 416 {
 417         watchdog_disable(cpu);
 418 }
 419
 420 static int watchdog_should_run(unsigned int cpu)
 421 {
 422         return __this_cpu_read(hrtimer_interrupts) !=
 423                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 424 }
 425
 426 /*
 427  * The watchdog thread function - touches the timestamp.
 428  *
 429  * It only runs once every sample_period seconds (4 seconds by
 430  * default) to reset the softlockup timestamp. If this gets delayed
 431  * for more than 2*watchdog_thresh seconds then the debug-printout
 432  * triggers in watchdog_timer_fn().
 433  */
 434 static void watchdog(unsigned int cpu)
 435 {
 436         __this_cpu_write(soft_lockup_hrtimer_cnt,
 437                          __this_cpu_read(hrtimer_interrupts));
 438         __touch_watchdog();
 439 }
 440
 441 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 442 /*
 443  * People like the simple clean cpu node info on boot.
 444  * Reduce the watchdog noise by only printing messages
 445  * that are different from what cpu0 displayed.
 446  */
 447 static unsigned long cpu0_err;
 448
 449 static int watchdog_nmi_enable(unsigned int cpu)
 450 {
 451         struct perf_event_attr *wd_attr;
 452         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 453
 454         /* is it already setup and enabled? */
 455         if (event && event->state > PERF_EVENT_STATE_OFF)
 456                 goto out;
 457
 458         /* it is setup but not enabled */
 459         if (event != NULL)
 460                 goto out_enable;
 461
 462         wd_attr = &wd_hw_attr;
 463         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 464
 465         /* Try to register using hardware perf events */
 466         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 467
 468         /* save cpu0 error for future comparision */
 469         if (cpu == 0 && IS_ERR(event))
 470                 cpu0_err = PTR_ERR(event);
 471
 472         if (!IS_ERR(event)) {
 473                 /* only print for cpu0 or different than cpu0 */
 474                 if (cpu == 0 || cpu0_err)
 475                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 476                 goto out_save;
 477         }
 478
 479         /* skip displaying the same error again */
 480         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 481                 return PTR_ERR(event);
 482
 483         /* vary the KERN level based on the returned errno */
 484         if (PTR_ERR(event) == -EOPNOTSUPP)
 485                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 486         else if (PTR_ERR(event) == -ENOENT)
 487                 pr_warning("disabled (cpu%i): hardware events not enabled\n",
 488                          cpu);
 489         else
 490                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 491                         cpu, PTR_ERR(event));
 492         return PTR_ERR(event);
 493
 494         /* success path */
 495 out_save:
 496         per_cpu(watchdog_ev, cpu) = event;
 497 out_enable:
 498         perf_event_enable(per_cpu(watchdog_ev, cpu));
 499 out:
 500         return 0;
 501 }
 502
 503 static void watchdog_nmi_disable(unsigned int cpu)
 504 {
 505         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 506
 507         if (event) {
 508                 perf_event_disable(event);
 509                 per_cpu(watchdog_ev, cpu) = NULL;
 510
 511                 /* should be in cleanup, but blocks oprofile */
 512                 perf_event_release_kernel(event);
 513         }
 514         return;
 515 }
 516 #else
 517 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 518 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 519 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 520
 521 static struct smp_hotplug_thread watchdog_threads = {
 522         .store                  = &softlockup_watchdog,
 523         .thread_should_run      = watchdog_should_run,
 524         .thread_fn              = watchdog,
 525         .thread_comm            = "watchdog/%u",
 526         .setup                  = watchdog_enable,
 527         .cleanup                = watchdog_cleanup,
 528         .park                   = watchdog_disable,
 529         .unpark                 = watchdog_enable,
 530 };
 531
 532 static void restart_watchdog_hrtimer(void *info)
 533 {
 534         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 535         int ret;
 536
 537         /*
 538          * No need to cancel and restart hrtimer if it is currently executing
 539          * because it will reprogram itself with the new period now.
 540          * We should never see it unqueued here because we are running per-cpu
 541          * with interrupts disabled.
 542          */
 543         ret = hrtimer_try_to_cancel(hrtimer);
 544         if (ret == 1)
 545                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 546                                 HRTIMER_MODE_REL_PINNED);
 547 }
 548
 549 static void update_timers(int cpu)
 550 {
 551         /*
 552          * Make sure that perf event counter will adopt to a new
 553          * sampling period. Updating the sampling period directly would
 554          * be much nicer but we do not have an API for that now so
 555          * let's use a big hammer.
 556          * Hrtimer will adopt the new period on the next tick but this
 557          * might be late already so we have to restart the timer as well.
 558          */
 559         watchdog_nmi_disable(cpu);
 560         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 561         watchdog_nmi_enable(cpu);
 562 }
 563
 564 static void update_timers_all_cpus(void)
 565 {
 566         int cpu;
 567
 568         get_online_cpus();
 569         for_each_online_cpu(cpu)
 570                 update_timers(cpu);
 571         put_online_cpus();
 572 }
 573
 574 static int watchdog_enable_all_cpus(bool sample_period_changed)
 575 {
 576         int err = 0;
 577
 578         if (!watchdog_running) {
 579                 err = smpboot_register_percpu_thread(&watchdog_threads);
 580                 if (err)
 581                         pr_err("Failed to create watchdog threads, disabled\n");
 582                 else
 583                         watchdog_running = 1;
 584         } else if (sample_period_changed) {
 585                 update_timers_all_cpus();
 586         }
 587
 588         return err;
 589 }
 590
 591 /* prepare/enable/disable routines */
 592 /* sysctl functions */
 593 #ifdef CONFIG_SYSCTL
 594 static void watchdog_disable_all_cpus(void)
 595 {
 596         if (watchdog_running) {
 597                 watchdog_running = 0;
 598                 smpboot_unregister_percpu_thread(&watchdog_threads);
 599         }
 600 }
 601
 602 /*
 603  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 604  */
 605
 606 int proc_dowatchdog(struct ctl_table *table, int write,
 607                     void __user *buffer, size_t *lenp, loff_t *ppos)
 608 {
 609         int err, old_thresh, old_enabled;
 610         static DEFINE_MUTEX(watchdog_proc_mutex);
 611
 612         mutex_lock(&watchdog_proc_mutex);
 613         old_thresh = ACCESS_ONCE(watchdog_thresh);
 614         old_enabled = ACCESS_ONCE(watchdog_user_enabled);
 615
 616         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 617         if (err || !write)
 618                 goto out;
 619
 620         set_sample_period();
 621         /*
 622          * Watchdog threads shouldn't be enabled if they are
 623          * disabled. The 'watchdog_running' variable check in
 624          * watchdog_*_all_cpus() function takes care of this.
 625          */
 626         if (watchdog_user_enabled && watchdog_thresh)
 627                 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
 628         else
 629                 watchdog_disable_all_cpus();
 630
 631         /* Restore old values on failure */
 632         if (err) {
 633                 watchdog_thresh = old_thresh;
 634                 watchdog_user_enabled = old_enabled;
 635         }
 636 out:
 637         mutex_unlock(&watchdog_proc_mutex);
 638         return err;
 639 }
 640 #endif /* CONFIG_SYSCTL */
 641
 642 void __init lockup_detector_init(void)
 643 {
 644         set_sample_period();
 645
 646         if (watchdog_user_enabled)
 647                 watchdog_enable_all_cpus(false);
 648 }