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