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