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