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",
 264                               this_cpu);
 265                 else
 266                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
 267                              this_cpu);
 268
 269                 __this_cpu_write(hard_watchdog_warn, true);
 270                 return;
 271         }
 272
 273         __this_cpu_write(hard_watchdog_warn, false);
 274         return;
 275 }
 276 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 277
 278 static void watchdog_interrupt_count(void)
 279 {
 280         __this_cpu_inc(hrtimer_interrupts);
 281 }
 282
 283 static int watchdog_nmi_enable(unsigned int cpu);
 284 static void watchdog_nmi_disable(unsigned int cpu);
 285
 286 /* watchdog kicker functions */
 287 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 288 {
 289         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 290         struct pt_regs *regs = get_irq_regs();
 291         int duration;
 292         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 293
 294         /* kick the hardlockup detector */
 295         watchdog_interrupt_count();
 296
 297         /* kick the softlockup detector */
 298         wake_up_process(__this_cpu_read(softlockup_watchdog));
 299
 300         /* .. and repeat */
 301         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 302
 303         if (touch_ts == 0) {
 304                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 305                         /*
 306                          * If the time stamp was touched atomically
 307                          * make sure the scheduler tick is up to date.
 308                          */
 309                         __this_cpu_write(softlockup_touch_sync, false);
 310                         sched_clock_tick();
 311                 }
 312
 313                 /* Clear the guest paused flag on watchdog reset */
 314                 kvm_check_and_clear_guest_paused();
 315                 __touch_watchdog();
 316                 return HRTIMER_RESTART;
 317         }
 318
 319         /* check for a softlockup
 320          * This is done by making sure a high priority task is
 321          * being scheduled.  The task touches the watchdog to
 322          * indicate it is getting cpu time.  If it hasn't then
 323          * this is a good indication some task is hogging the cpu
 324          */
 325         duration = is_softlockup(touch_ts);
 326         if (unlikely(duration)) {
 327                 /*
 328                  * If a virtual machine is stopped by the host it can look to
 329                  * the watchdog like a soft lockup, check to see if the host
 330                  * stopped the vm before we issue the warning
 331                  */
 332                 if (kvm_check_and_clear_guest_paused())
 333                         return HRTIMER_RESTART;
 334
 335                 /* only warn once */
 336                 if (__this_cpu_read(soft_watchdog_warn) == true)
 337                         return HRTIMER_RESTART;
 338
 339                 if (softlockup_all_cpu_backtrace) {
 340                         /* Prevent multiple soft-lockup reports if one cpu is already
 341                          * engaged in dumping cpu back traces
 342                          */
 343                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 344                                 /* Someone else will report us. Let's give up */
 345                                 __this_cpu_write(soft_watchdog_warn, true);
 346                                 return HRTIMER_RESTART;
 347                         }
 348                 }
 349
 350                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 351                         smp_processor_id(), duration,
 352                         current->comm, task_pid_nr(current));
 353                 print_modules();
 354                 print_irqtrace_events(current);
 355                 if (regs)
 356                         show_regs(regs);
 357                 else
 358                         dump_stack();
 359
 360                 if (softlockup_all_cpu_backtrace) {
 361                         /* Avoid generating two back traces for current
 362                          * given that one is already made above
 363                          */
 364                         trigger_allbutself_cpu_backtrace();
 365
 366                         clear_bit(0, &soft_lockup_nmi_warn);
 367                         /* Barrier to sync with other cpus */
 368                         smp_mb__after_atomic();
 369                 }
 370
 371                 if (softlockup_panic)
 372                         panic("softlockup: hung tasks");
 373                 __this_cpu_write(soft_watchdog_warn, true);
 374         } else
 375                 __this_cpu_write(soft_watchdog_warn, false);
 376
 377         return HRTIMER_RESTART;
 378 }
 379
 380 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 381 {
 382         struct sched_param param = { .sched_priority = prio };
 383
 384         sched_setscheduler(current, policy, &param);
 385 }
 386
 387 static void watchdog_enable(unsigned int cpu)
 388 {
 389         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 390
 391         /* kick off the timer for the hardlockup detector */
 392         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 393         hrtimer->function = watchdog_timer_fn;
 394
 395         /* Enable the perf event */
 396         watchdog_nmi_enable(cpu);
 397
 398         /* done here because hrtimer_start can only pin to smp_processor_id() */
 399         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 400                       HRTIMER_MODE_REL_PINNED);
 401
 402         /* initialize timestamp */
 403         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 404         __touch_watchdog();
 405 }
 406
 407 static void watchdog_disable(unsigned int cpu)
 408 {
 409         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 410
 411         watchdog_set_prio(SCHED_NORMAL, 0);
 412         hrtimer_cancel(hrtimer);
 413         /* disable the perf event */
 414         watchdog_nmi_disable(cpu);
 415 }
 416
 417 static void watchdog_cleanup(unsigned int cpu, bool online)
 418 {
 419         watchdog_disable(cpu);
 420 }
 421
 422 static int watchdog_should_run(unsigned int cpu)
 423 {
 424         return __this_cpu_read(hrtimer_interrupts) !=
 425                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 426 }
 427
 428 /*
 429  * The watchdog thread function - touches the timestamp.
 430  *
 431  * It only runs once every sample_period seconds (4 seconds by
 432  * default) to reset the softlockup timestamp. If this gets delayed
 433  * for more than 2*watchdog_thresh seconds then the debug-printout
 434  * triggers in watchdog_timer_fn().
 435  */
 436 static void watchdog(unsigned int cpu)
 437 {
 438         __this_cpu_write(soft_lockup_hrtimer_cnt,
 439                          __this_cpu_read(hrtimer_interrupts));
 440         __touch_watchdog();
 441 }
 442
 443 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 444 /*
 445  * People like the simple clean cpu node info on boot.
 446  * Reduce the watchdog noise by only printing messages
 447  * that are different from what cpu0 displayed.
 448  */
 449 static unsigned long cpu0_err;
 450
 451 static int watchdog_nmi_enable(unsigned int cpu)
 452 {
 453         struct perf_event_attr *wd_attr;
 454         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 455
 456         /* is it already setup and enabled? */
 457         if (event && event->state > PERF_EVENT_STATE_OFF)
 458                 goto out;
 459
 460         /* it is setup but not enabled */
 461         if (event != NULL)
 462                 goto out_enable;
 463
 464         wd_attr = &wd_hw_attr;
 465         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 466
 467         /* Try to register using hardware perf events */
 468         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 469
 470         /* save cpu0 error for future comparision */
 471         if (cpu == 0 && IS_ERR(event))
 472                 cpu0_err = PTR_ERR(event);
 473
 474         if (!IS_ERR(event)) {
 475                 /* only print for cpu0 or different than cpu0 */
 476                 if (cpu == 0 || cpu0_err)
 477                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 478                 goto out_save;
 479         }
 480
 481         /* skip displaying the same error again */
 482         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 483                 return PTR_ERR(event);
 484
 485         /* vary the KERN level based on the returned errno */
 486         if (PTR_ERR(event) == -EOPNOTSUPP)
 487                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 488         else if (PTR_ERR(event) == -ENOENT)
 489                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 490                          cpu);
 491         else
 492                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 493                         cpu, PTR_ERR(event));
 494         return PTR_ERR(event);
 495
 496         /* success path */
 497 out_save:
 498         per_cpu(watchdog_ev, cpu) = event;
 499 out_enable:
 500         perf_event_enable(per_cpu(watchdog_ev, cpu));
 501 out:
 502         return 0;
 503 }
 504
 505 static void watchdog_nmi_disable(unsigned int cpu)
 506 {
 507         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 508
 509         if (event) {
 510                 perf_event_disable(event);
 511                 per_cpu(watchdog_ev, cpu) = NULL;
 512
 513                 /* should be in cleanup, but blocks oprofile */
 514                 perf_event_release_kernel(event);
 515         }
 516         return;
 517 }
 518 #else
 519 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 520 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 521 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 522
 523 static struct smp_hotplug_thread watchdog_threads = {
 524         .store                  = &softlockup_watchdog,
 525         .thread_should_run      = watchdog_should_run,
 526         .thread_fn              = watchdog,
 527         .thread_comm            = "watchdog/%u",
 528         .setup                  = watchdog_enable,
 529         .cleanup                = watchdog_cleanup,
 530         .park                   = watchdog_disable,
 531         .unpark                 = watchdog_enable,
 532 };
 533
 534 static void restart_watchdog_hrtimer(void *info)
 535 {
 536         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 537         int ret;
 538
 539         /*
 540          * No need to cancel and restart hrtimer if it is currently executing
 541          * because it will reprogram itself with the new period now.
 542          * We should never see it unqueued here because we are running per-cpu
 543          * with interrupts disabled.
 544          */
 545         ret = hrtimer_try_to_cancel(hrtimer);
 546         if (ret == 1)
 547                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 548                                 HRTIMER_MODE_REL_PINNED);
 549 }
 550
 551 static void update_timers(int cpu)
 552 {
 553         /*
 554          * Make sure that perf event counter will adopt to a new
 555          * sampling period. Updating the sampling period directly would
 556          * be much nicer but we do not have an API for that now so
 557          * let's use a big hammer.
 558          * Hrtimer will adopt the new period on the next tick but this
 559          * might be late already so we have to restart the timer as well.
 560          */
 561         watchdog_nmi_disable(cpu);
 562         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 563         watchdog_nmi_enable(cpu);
 564 }
 565
 566 static void update_timers_all_cpus(void)
 567 {
 568         int cpu;
 569
 570         get_online_cpus();
 571         for_each_online_cpu(cpu)
 572                 update_timers(cpu);
 573         put_online_cpus();
 574 }
 575
 576 static int watchdog_enable_all_cpus(bool sample_period_changed)
 577 {
 578         int err = 0;
 579
 580         if (!watchdog_running) {
 581                 err = smpboot_register_percpu_thread(&watchdog_threads);
 582                 if (err)
 583                         pr_err("Failed to create watchdog threads, disabled\n");
 584                 else
 585                         watchdog_running = 1;
 586         } else if (sample_period_changed) {
 587                 update_timers_all_cpus();
 588         }
 589
 590         return err;
 591 }
 592
 593 /* prepare/enable/disable routines */
 594 /* sysctl functions */
 595 #ifdef CONFIG_SYSCTL
 596 static void watchdog_disable_all_cpus(void)
 597 {
 598         if (watchdog_running) {
 599                 watchdog_running = 0;
 600                 smpboot_unregister_percpu_thread(&watchdog_threads);
 601         }
 602 }
 603
 604 /*
 605  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 606  */
 607
 608 int proc_dowatchdog(struct ctl_table *table, int write,
 609                     void __user *buffer, size_t *lenp, loff_t *ppos)
 610 {
 611         int err, old_thresh, old_enabled;
 612         static DEFINE_MUTEX(watchdog_proc_mutex);
 613
 614         mutex_lock(&watchdog_proc_mutex);
 615         old_thresh = ACCESS_ONCE(watchdog_thresh);
 616         old_enabled = ACCESS_ONCE(watchdog_user_enabled);
 617
 618         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 619         if (err || !write)
 620                 goto out;
 621
 622         set_sample_period();
 623         /*
 624          * Watchdog threads shouldn't be enabled if they are
 625          * disabled. The 'watchdog_running' variable check in
 626          * watchdog_*_all_cpus() function takes care of this.
 627          */
 628         if (watchdog_user_enabled && watchdog_thresh)
 629                 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
 630         else
 631                 watchdog_disable_all_cpus();
 632
 633         /* Restore old values on failure */
 634         if (err) {
 635                 watchdog_thresh = old_thresh;
 636                 watchdog_user_enabled = old_enabled;
 637         }
 638 out:
 639         mutex_unlock(&watchdog_proc_mutex);
 640         return err;
 641 }
 642 #endif /* CONFIG_SYSCTL */
 643
 644 void __init lockup_detector_init(void)
 645 {
 646         set_sample_period();
 647
 648         if (watchdog_user_enabled)
 649                 watchdog_enable_all_cpus(false);
 650 }