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         /*
 142          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 143          * gets zeroed here, so use the raw_ operation.
 144          */
 145         raw_cpu_write(watchdog_touch_ts, 0);
 146 }
 147 EXPORT_SYMBOL(touch_softlockup_watchdog);
 148
 149 void touch_all_softlockup_watchdogs(void)
 150 {
 151         int cpu;
 152
 153         /*
 154          * this is done lockless
 155          * do we care if a 0 races with a timestamp?
 156          * all it means is the softlock check starts one cycle later
 157          */
 158         for_each_online_cpu(cpu)
 159                 per_cpu(watchdog_touch_ts, cpu) = 0;
 160 }
 161
 162 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 163 void touch_nmi_watchdog(void)
 164 {
 165         /*
 166          * Using __raw here because some code paths have
 167          * preemption enabled.  If preemption is enabled
 168          * then interrupts should be enabled too, in which
 169          * case we shouldn't have to worry about the watchdog
 170          * going off.
 171          */
 172         __raw_get_cpu_var(watchdog_nmi_touch) = true;
 173         touch_softlockup_watchdog();
 174 }
 175 EXPORT_SYMBOL(touch_nmi_watchdog);
 176
 177 #endif
 178
 179 void touch_softlockup_watchdog_sync(void)
 180 {
 181         __raw_get_cpu_var(softlockup_touch_sync) = true;
 182         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 183 }
 184
 185 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 186 /* watchdog detector functions */
 187 static int is_hardlockup(void)
 188 {
 189         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 190
 191         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 192                 return 1;
 193
 194         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 195         return 0;
 196 }
 197 #endif
 198
 199 static int is_softlockup(unsigned long touch_ts)
 200 {
 201         unsigned long now = get_timestamp();
 202
 203         /* Warn about unreasonable delays: */
 204         if (time_after(now, touch_ts + get_softlockup_thresh()))
 205                 return now - touch_ts;
 206
 207         return 0;
 208 }
 209
 210 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 211
 212 static struct perf_event_attr wd_hw_attr = {
 213         .type           = PERF_TYPE_HARDWARE,
 214         .config         = PERF_COUNT_HW_CPU_CYCLES,
 215         .size           = sizeof(struct perf_event_attr),
 216         .pinned         = 1,
 217         .disabled       = 1,
 218 };
 219
 220 /* Callback function for perf event subsystem */
 221 static void watchdog_overflow_callback(struct perf_event *event,
 222                  struct perf_sample_data *data,
 223                  struct pt_regs *regs)
 224 {
 225         /* Ensure the watchdog never gets throttled */
 226         event->hw.interrupts = 0;
 227
 228         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 229                 __this_cpu_write(watchdog_nmi_touch, false);
 230                 return;
 231         }
 232
 233         /* check for a hardlockup
 234          * This is done by making sure our timer interrupt
 235          * is incrementing.  The timer interrupt should have
 236          * fired multiple times before we overflow'd.  If it hasn't
 237          * then this is a good indication the cpu is stuck
 238          */
 239         if (is_hardlockup()) {
 240                 int this_cpu = smp_processor_id();
 241
 242                 /* only print hardlockups once */
 243                 if (__this_cpu_read(hard_watchdog_warn) == true)
 244                         return;
 245
 246                 if (hardlockup_panic)
 247                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 248                 else
 249                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 250
 251                 __this_cpu_write(hard_watchdog_warn, true);
 252                 return;
 253         }
 254
 255         __this_cpu_write(hard_watchdog_warn, false);
 256         return;
 257 }
 258 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 259
 260 static void watchdog_interrupt_count(void)
 261 {
 262         __this_cpu_inc(hrtimer_interrupts);
 263 }
 264
 265 static int watchdog_nmi_enable(unsigned int cpu);
 266 static void watchdog_nmi_disable(unsigned int cpu);
 267
 268 /* watchdog kicker functions */
 269 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 270 {
 271         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 272         struct pt_regs *regs = get_irq_regs();
 273         int duration;
 274
 275         /* kick the hardlockup detector */
 276         watchdog_interrupt_count();
 277
 278         /* kick the softlockup detector */
 279         wake_up_process(__this_cpu_read(softlockup_watchdog));
 280
 281         /* .. and repeat */
 282         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 283
 284         if (touch_ts == 0) {
 285                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 286                         /*
 287                          * If the time stamp was touched atomically
 288                          * make sure the scheduler tick is up to date.
 289                          */
 290                         __this_cpu_write(softlockup_touch_sync, false);
 291                         sched_clock_tick();
 292                 }
 293
 294                 /* Clear the guest paused flag on watchdog reset */
 295                 kvm_check_and_clear_guest_paused();
 296                 __touch_watchdog();
 297                 return HRTIMER_RESTART;
 298         }
 299
 300         /* check for a softlockup
 301          * This is done by making sure a high priority task is
 302          * being scheduled.  The task touches the watchdog to
 303          * indicate it is getting cpu time.  If it hasn't then
 304          * this is a good indication some task is hogging the cpu
 305          */
 306         duration = is_softlockup(touch_ts);
 307         if (unlikely(duration)) {
 308                 /*
 309                  * If a virtual machine is stopped by the host it can look to
 310                  * the watchdog like a soft lockup, check to see if the host
 311                  * stopped the vm before we issue the warning
 312                  */
 313                 if (kvm_check_and_clear_guest_paused())
 314                         return HRTIMER_RESTART;
 315
 316                 /* only warn once */
 317                 if (__this_cpu_read(soft_watchdog_warn) == true)
 318                         return HRTIMER_RESTART;
 319
 320                 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 321                         smp_processor_id(), duration,
 322                         current->comm, task_pid_nr(current));
 323                 print_modules();
 324                 print_irqtrace_events(current);
 325                 if (regs)
 326                         show_regs(regs);
 327                 else
 328                         dump_stack();
 329
 330                 if (softlockup_panic)
 331                         panic("softlockup: hung tasks");
 332                 __this_cpu_write(soft_watchdog_warn, true);
 333         } else
 334                 __this_cpu_write(soft_watchdog_warn, false);
 335
 336         return HRTIMER_RESTART;
 337 }
 338
 339 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 340 {
 341         struct sched_param param = { .sched_priority = prio };
 342
 343         sched_setscheduler(current, policy, &param);
 344 }
 345
 346 static void watchdog_enable(unsigned int cpu)
 347 {
 348         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 349
 350         /* kick off the timer for the hardlockup detector */
 351         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 352         hrtimer->function = watchdog_timer_fn;
 353
 354         /* Enable the perf event */
 355         watchdog_nmi_enable(cpu);
 356
 357         /* done here because hrtimer_start can only pin to smp_processor_id() */
 358         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 359                       HRTIMER_MODE_REL_PINNED);
 360
 361         /* initialize timestamp */
 362         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 363         __touch_watchdog();
 364 }
 365
 366 static void watchdog_disable(unsigned int cpu)
 367 {
 368         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 369
 370         watchdog_set_prio(SCHED_NORMAL, 0);
 371         hrtimer_cancel(hrtimer);
 372         /* disable the perf event */
 373         watchdog_nmi_disable(cpu);
 374 }
 375
 376 static void watchdog_cleanup(unsigned int cpu, bool online)
 377 {
 378         watchdog_disable(cpu);
 379 }
 380
 381 static int watchdog_should_run(unsigned int cpu)
 382 {
 383         return __this_cpu_read(hrtimer_interrupts) !=
 384                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 385 }
 386
 387 /*
 388  * The watchdog thread function - touches the timestamp.
 389  *
 390  * It only runs once every sample_period seconds (4 seconds by
 391  * default) to reset the softlockup timestamp. If this gets delayed
 392  * for more than 2*watchdog_thresh seconds then the debug-printout
 393  * triggers in watchdog_timer_fn().
 394  */
 395 static void watchdog(unsigned int cpu)
 396 {
 397         __this_cpu_write(soft_lockup_hrtimer_cnt,
 398                          __this_cpu_read(hrtimer_interrupts));
 399         __touch_watchdog();
 400 }
 401
 402 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 403 /*
 404  * People like the simple clean cpu node info on boot.
 405  * Reduce the watchdog noise by only printing messages
 406  * that are different from what cpu0 displayed.
 407  */
 408 static unsigned long cpu0_err;
 409
 410 static int watchdog_nmi_enable(unsigned int cpu)
 411 {
 412         struct perf_event_attr *wd_attr;
 413         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 414
 415         /* is it already setup and enabled? */
 416         if (event && event->state > PERF_EVENT_STATE_OFF)
 417                 goto out;
 418
 419         /* it is setup but not enabled */
 420         if (event != NULL)
 421                 goto out_enable;
 422
 423         wd_attr = &wd_hw_attr;
 424         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 425
 426         /* Try to register using hardware perf events */
 427         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 428
 429         /* save cpu0 error for future comparision */
 430         if (cpu == 0 && IS_ERR(event))
 431                 cpu0_err = PTR_ERR(event);
 432
 433         if (!IS_ERR(event)) {
 434                 /* only print for cpu0 or different than cpu0 */
 435                 if (cpu == 0 || cpu0_err)
 436                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 437                 goto out_save;
 438         }
 439
 440         /* skip displaying the same error again */
 441         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 442                 return PTR_ERR(event);
 443
 444         /* vary the KERN level based on the returned errno */
 445         if (PTR_ERR(event) == -EOPNOTSUPP)
 446                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 447         else if (PTR_ERR(event) == -ENOENT)
 448                 pr_warning("disabled (cpu%i): hardware events not enabled\n",
 449                          cpu);
 450         else
 451                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 452                         cpu, PTR_ERR(event));
 453         return PTR_ERR(event);
 454
 455         /* success path */
 456 out_save:
 457         per_cpu(watchdog_ev, cpu) = event;
 458 out_enable:
 459         perf_event_enable(per_cpu(watchdog_ev, cpu));
 460 out:
 461         return 0;
 462 }
 463
 464 static void watchdog_nmi_disable(unsigned int cpu)
 465 {
 466         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 467
 468         if (event) {
 469                 perf_event_disable(event);
 470                 per_cpu(watchdog_ev, cpu) = NULL;
 471
 472                 /* should be in cleanup, but blocks oprofile */
 473                 perf_event_release_kernel(event);
 474         }
 475         return;
 476 }
 477 #else
 478 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 479 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 480 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 481
 482 static struct smp_hotplug_thread watchdog_threads = {
 483         .store                  = &softlockup_watchdog,
 484         .thread_should_run      = watchdog_should_run,
 485         .thread_fn              = watchdog,
 486         .thread_comm            = "watchdog/%u",
 487         .setup                  = watchdog_enable,
 488         .cleanup                = watchdog_cleanup,
 489         .park                   = watchdog_disable,
 490         .unpark                 = watchdog_enable,
 491 };
 492
 493 static void restart_watchdog_hrtimer(void *info)
 494 {
 495         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 496         int ret;
 497
 498         /*
 499          * No need to cancel and restart hrtimer if it is currently executing
 500          * because it will reprogram itself with the new period now.
 501          * We should never see it unqueued here because we are running per-cpu
 502          * with interrupts disabled.
 503          */
 504         ret = hrtimer_try_to_cancel(hrtimer);
 505         if (ret == 1)
 506                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 507                                 HRTIMER_MODE_REL_PINNED);
 508 }
 509
 510 static void update_timers(int cpu)
 511 {
 512         /*
 513          * Make sure that perf event counter will adopt to a new
 514          * sampling period. Updating the sampling period directly would
 515          * be much nicer but we do not have an API for that now so
 516          * let's use a big hammer.
 517          * Hrtimer will adopt the new period on the next tick but this
 518          * might be late already so we have to restart the timer as well.
 519          */
 520         watchdog_nmi_disable(cpu);
 521         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 522         watchdog_nmi_enable(cpu);
 523 }
 524
 525 static void update_timers_all_cpus(void)
 526 {
 527         int cpu;
 528
 529         get_online_cpus();
 530         preempt_disable();
 531         for_each_online_cpu(cpu)
 532                 update_timers(cpu);
 533         preempt_enable();
 534         put_online_cpus();
 535 }
 536
 537 static int watchdog_enable_all_cpus(bool sample_period_changed)
 538 {
 539         int err = 0;
 540
 541         if (!watchdog_running) {
 542                 err = smpboot_register_percpu_thread(&watchdog_threads);
 543                 if (err)
 544                         pr_err("Failed to create watchdog threads, disabled\n");
 545                 else
 546                         watchdog_running = 1;
 547         } else if (sample_period_changed) {
 548                 update_timers_all_cpus();
 549         }
 550
 551         return err;
 552 }
 553
 554 /* prepare/enable/disable routines */
 555 /* sysctl functions */
 556 #ifdef CONFIG_SYSCTL
 557 static void watchdog_disable_all_cpus(void)
 558 {
 559         if (watchdog_running) {
 560                 watchdog_running = 0;
 561                 smpboot_unregister_percpu_thread(&watchdog_threads);
 562         }
 563 }
 564
 565 /*
 566  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 567  */
 568
 569 int proc_dowatchdog(struct ctl_table *table, int write,
 570                     void __user *buffer, size_t *lenp, loff_t *ppos)
 571 {
 572         int err, old_thresh, old_enabled;
 573         static DEFINE_MUTEX(watchdog_proc_mutex);
 574
 575         mutex_lock(&watchdog_proc_mutex);
 576         old_thresh = ACCESS_ONCE(watchdog_thresh);
 577         old_enabled = ACCESS_ONCE(watchdog_user_enabled);
 578
 579         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 580         if (err || !write)
 581                 goto out;
 582
 583         set_sample_period();
 584         /*
 585          * Watchdog threads shouldn't be enabled if they are
 586          * disabled. The 'watchdog_running' variable check in
 587          * watchdog_*_all_cpus() function takes care of this.
 588          */
 589         if (watchdog_user_enabled && watchdog_thresh)
 590                 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
 591         else
 592                 watchdog_disable_all_cpus();
 593
 594         /* Restore old values on failure */
 595         if (err) {
 596                 watchdog_thresh = old_thresh;
 597                 watchdog_user_enabled = old_enabled;
 598         }
 599 out:
 600         mutex_unlock(&watchdog_proc_mutex);
 601         return err;
 602 }
 603 #endif /* CONFIG_SYSCTL */
 604
 605 void __init lockup_detector_init(void)
 606 {
 607         set_sample_period();
 608
 609         if (watchdog_user_enabled)
 610                 watchdog_enable_all_cpus(false);
 611 }