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_enabled = 1;
  33 int __read_mostly watchdog_thresh = 10;
  34 static int __read_mostly watchdog_disabled;
  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_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_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_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(int this_cpu)
 117 {
 118         return cpu_clock(this_cpu) >> 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         int this_cpu = smp_processor_id();
 137
 138         __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
 139 }
 140
 141 void touch_softlockup_watchdog(void)
 142 {
 143         __this_cpu_write(watchdog_touch_ts, 0);
 144 }
 145 EXPORT_SYMBOL(touch_softlockup_watchdog);
 146
 147 void touch_all_softlockup_watchdogs(void)
 148 {
 149         int cpu;
 150
 151         /*
 152          * this is done lockless
 153          * do we care if a 0 races with a timestamp?
 154          * all it means is the softlock check starts one cycle later
 155          */
 156         for_each_online_cpu(cpu)
 157                 per_cpu(watchdog_touch_ts, cpu) = 0;
 158 }
 159
 160 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 161 void touch_nmi_watchdog(void)
 162 {
 163         if (watchdog_enabled) {
 164                 unsigned cpu;
 165
 166                 for_each_present_cpu(cpu) {
 167                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
 168                                 per_cpu(watchdog_nmi_touch, cpu) = true;
 169                 }
 170         }
 171         touch_softlockup_watchdog();
 172 }
 173 EXPORT_SYMBOL(touch_nmi_watchdog);
 174
 175 #endif
 176
 177 void touch_softlockup_watchdog_sync(void)
 178 {
 179         __raw_get_cpu_var(softlockup_touch_sync) = true;
 180         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 181 }
 182
 183 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 184 /* watchdog detector functions */
 185 static int is_hardlockup(void)
 186 {
 187         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 188
 189         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 190                 return 1;
 191
 192         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 193         return 0;
 194 }
 195 #endif
 196
 197 static int is_softlockup(unsigned long touch_ts)
 198 {
 199         unsigned long now = get_timestamp(smp_processor_id());
 200
 201         /* Warn about unreasonable delays: */
 202         if (time_after(now, touch_ts + get_softlockup_thresh()))
 203                 return now - touch_ts;
 204
 205         return 0;
 206 }
 207
 208 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 209
 210 static struct perf_event_attr wd_hw_attr = {
 211         .type           = PERF_TYPE_HARDWARE,
 212         .config         = PERF_COUNT_HW_CPU_CYCLES,
 213         .size           = sizeof(struct perf_event_attr),
 214         .pinned         = 1,
 215         .disabled       = 1,
 216 };
 217
 218 /* Callback function for perf event subsystem */
 219 static void watchdog_overflow_callback(struct perf_event *event,
 220                  struct perf_sample_data *data,
 221                  struct pt_regs *regs)
 222 {
 223         /* Ensure the watchdog never gets throttled */
 224         event->hw.interrupts = 0;
 225
 226         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 227                 __this_cpu_write(watchdog_nmi_touch, false);
 228                 return;
 229         }
 230
 231         /* check for a hardlockup
 232          * This is done by making sure our timer interrupt
 233          * is incrementing.  The timer interrupt should have
 234          * fired multiple times before we overflow'd.  If it hasn't
 235          * then this is a good indication the cpu is stuck
 236          */
 237         if (is_hardlockup()) {
 238                 int this_cpu = smp_processor_id();
 239
 240                 /* only print hardlockups once */
 241                 if (__this_cpu_read(hard_watchdog_warn) == true)
 242                         return;
 243
 244                 if (hardlockup_panic)
 245                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 246                 else
 247                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 248
 249                 __this_cpu_write(hard_watchdog_warn, true);
 250                 return;
 251         }
 252
 253         __this_cpu_write(hard_watchdog_warn, false);
 254         return;
 255 }
 256 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 257
 258 static void watchdog_interrupt_count(void)
 259 {
 260         __this_cpu_inc(hrtimer_interrupts);
 261 }
 262
 263 static int watchdog_nmi_enable(unsigned int cpu);
 264 static void watchdog_nmi_disable(unsigned int cpu);
 265
 266 /* watchdog kicker functions */
 267 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 268 {
 269         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 270         struct pt_regs *regs = get_irq_regs();
 271         int duration;
 272
 273         /* kick the hardlockup detector */
 274         watchdog_interrupt_count();
 275
 276         /* kick the softlockup detector */
 277         wake_up_process(__this_cpu_read(softlockup_watchdog));
 278
 279         /* .. and repeat */
 280         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 281
 282         if (touch_ts == 0) {
 283                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 284                         /*
 285                          * If the time stamp was touched atomically
 286                          * make sure the scheduler tick is up to date.
 287                          */
 288                         __this_cpu_write(softlockup_touch_sync, false);
 289                         sched_clock_tick();
 290                 }
 291
 292                 /* Clear the guest paused flag on watchdog reset */
 293                 kvm_check_and_clear_guest_paused();
 294                 __touch_watchdog();
 295                 return HRTIMER_RESTART;
 296         }
 297
 298         /* check for a softlockup
 299          * This is done by making sure a high priority task is
 300          * being scheduled.  The task touches the watchdog to
 301          * indicate it is getting cpu time.  If it hasn't then
 302          * this is a good indication some task is hogging the cpu
 303          */
 304         duration = is_softlockup(touch_ts);
 305         if (unlikely(duration)) {
 306                 /*
 307                  * If a virtual machine is stopped by the host it can look to
 308                  * the watchdog like a soft lockup, check to see if the host
 309                  * stopped the vm before we issue the warning
 310                  */
 311                 if (kvm_check_and_clear_guest_paused())
 312                         return HRTIMER_RESTART;
 313
 314                 /* only warn once */
 315                 if (__this_cpu_read(soft_watchdog_warn) == true)
 316                         return HRTIMER_RESTART;
 317
 318                 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 319                         smp_processor_id(), duration,
 320                         current->comm, task_pid_nr(current));
 321                 print_modules();
 322                 print_irqtrace_events(current);
 323                 if (regs)
 324                         show_regs(regs);
 325                 else
 326                         dump_stack();
 327
 328                 if (softlockup_panic)
 329                         panic("softlockup: hung tasks");
 330                 __this_cpu_write(soft_watchdog_warn, true);
 331         } else
 332                 __this_cpu_write(soft_watchdog_warn, false);
 333
 334         return HRTIMER_RESTART;
 335 }
 336
 337 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 338 {
 339         struct sched_param param = { .sched_priority = prio };
 340
 341         sched_setscheduler(current, policy, &param);
 342 }
 343
 344 static void watchdog_enable(unsigned int cpu)
 345 {
 346         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 347
 348         /* kick off the timer for the hardlockup detector */
 349         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 350         hrtimer->function = watchdog_timer_fn;
 351
 352         if (!watchdog_enabled) {
 353                 kthread_park(current);
 354                 return;
 355         }
 356
 357         /* Enable the perf event */
 358         watchdog_nmi_enable(cpu);
 359
 360         /* done here because hrtimer_start can only pin to smp_processor_id() */
 361         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 362                       HRTIMER_MODE_REL_PINNED);
 363
 364         /* initialize timestamp */
 365         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 366         __touch_watchdog();
 367 }
 368
 369 static void watchdog_disable(unsigned int cpu)
 370 {
 371         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 372
 373         watchdog_set_prio(SCHED_NORMAL, 0);
 374         hrtimer_cancel(hrtimer);
 375         /* disable the perf event */
 376         watchdog_nmi_disable(cpu);
 377 }
 378
 379 static int watchdog_should_run(unsigned int cpu)
 380 {
 381         return __this_cpu_read(hrtimer_interrupts) !=
 382                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 383 }
 384
 385 /*
 386  * The watchdog thread function - touches the timestamp.
 387  *
 388  * It only runs once every sample_period seconds (4 seconds by
 389  * default) to reset the softlockup timestamp. If this gets delayed
 390  * for more than 2*watchdog_thresh seconds then the debug-printout
 391  * triggers in watchdog_timer_fn().
 392  */
 393 static void watchdog(unsigned int cpu)
 394 {
 395         __this_cpu_write(soft_lockup_hrtimer_cnt,
 396                          __this_cpu_read(hrtimer_interrupts));
 397         __touch_watchdog();
 398 }
 399
 400 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 401 /*
 402  * People like the simple clean cpu node info on boot.
 403  * Reduce the watchdog noise by only printing messages
 404  * that are different from what cpu0 displayed.
 405  */
 406 static unsigned long cpu0_err;
 407
 408 static int watchdog_nmi_enable(unsigned int cpu)
 409 {
 410         struct perf_event_attr *wd_attr;
 411         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 412
 413         /* is it already setup and enabled? */
 414         if (event && event->state > PERF_EVENT_STATE_OFF)
 415                 goto out;
 416
 417         /* it is setup but not enabled */
 418         if (event != NULL)
 419                 goto out_enable;
 420
 421         wd_attr = &wd_hw_attr;
 422         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 423
 424         /* Try to register using hardware perf events */
 425         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 426
 427         /* save cpu0 error for future comparision */
 428         if (cpu == 0 && IS_ERR(event))
 429                 cpu0_err = PTR_ERR(event);
 430
 431         if (!IS_ERR(event)) {
 432                 /* only print for cpu0 or different than cpu0 */
 433                 if (cpu == 0 || cpu0_err)
 434                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 435                 goto out_save;
 436         }
 437
 438         /* skip displaying the same error again */
 439         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 440                 return PTR_ERR(event);
 441
 442         /* vary the KERN level based on the returned errno */
 443         if (PTR_ERR(event) == -EOPNOTSUPP)
 444                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 445         else if (PTR_ERR(event) == -ENOENT)
 446                 pr_warning("disabled (cpu%i): hardware events not enabled\n",
 447                          cpu);
 448         else
 449                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 450                         cpu, PTR_ERR(event));
 451         return PTR_ERR(event);
 452
 453         /* success path */
 454 out_save:
 455         per_cpu(watchdog_ev, cpu) = event;
 456 out_enable:
 457         perf_event_enable(per_cpu(watchdog_ev, cpu));
 458 out:
 459         return 0;
 460 }
 461
 462 static void watchdog_nmi_disable(unsigned int cpu)
 463 {
 464         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 465
 466         if (event) {
 467                 perf_event_disable(event);
 468                 per_cpu(watchdog_ev, cpu) = NULL;
 469
 470                 /* should be in cleanup, but blocks oprofile */
 471                 perf_event_release_kernel(event);
 472         }
 473         return;
 474 }
 475 #else
 476 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 477 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 478 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 479
 480 /* prepare/enable/disable routines */
 481 /* sysctl functions */
 482 #ifdef CONFIG_SYSCTL
 483 static void watchdog_enable_all_cpus(void)
 484 {
 485         unsigned int cpu;
 486
 487         if (watchdog_disabled) {
 488                 watchdog_disabled = 0;
 489                 for_each_online_cpu(cpu)
 490                         kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 491         }
 492 }
 493
 494 static void watchdog_disable_all_cpus(void)
 495 {
 496         unsigned int cpu;
 497
 498         if (!watchdog_disabled) {
 499                 watchdog_disabled = 1;
 500                 for_each_online_cpu(cpu)
 501                         kthread_park(per_cpu(softlockup_watchdog, cpu));
 502         }
 503 }
 504
 505 /*
 506  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 507  */
 508
 509 int proc_dowatchdog(struct ctl_table *table, int write,
 510                     void __user *buffer, size_t *lenp, loff_t *ppos)
 511 {
 512         int ret;
 513
 514         if (watchdog_disabled < 0)
 515                 return -ENODEV;
 516
 517         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 518         if (ret || !write)
 519                 return ret;
 520
 521         set_sample_period();
 522         if (watchdog_enabled && watchdog_thresh)
 523                 watchdog_enable_all_cpus();
 524         else
 525                 watchdog_disable_all_cpus();
 526
 527         return ret;
 528 }
 529 #endif /* CONFIG_SYSCTL */
 530
 531 static struct smp_hotplug_thread watchdog_threads = {
 532         .store                  = &softlockup_watchdog,
 533         .thread_should_run      = watchdog_should_run,
 534         .thread_fn              = watchdog,
 535         .thread_comm            = "watchdog/%u",
 536         .setup                  = watchdog_enable,
 537         .park                   = watchdog_disable,
 538         .unpark                 = watchdog_enable,
 539 };
 540
 541 void __init lockup_detector_init(void)
 542 {
 543         set_sample_period();
 544         if (smpboot_register_percpu_thread(&watchdog_threads)) {
 545                 pr_err("Failed to create watchdog threads, disabled\n");
 546                 watchdog_disabled = -ENODEV;
 547         }
 548 }