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 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
  46 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  47 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  48 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  49 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  50 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  51 #endif
  52 static unsigned long soft_lockup_nmi_warn;
  53
  54 /* boot commands */
  55 /*
  56  * Should we panic when a soft-lockup or hard-lockup occurs:
  57  */
  58 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  59 static int hardlockup_panic =
  60                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  61
  62 static bool hardlockup_detector_enabled = true;
  63 /*
  64  * We may not want to enable hard lockup detection by default in all cases,
  65  * for example when running the kernel as a guest on a hypervisor. In these
  66  * cases this function can be called to disable hard lockup detection. This
  67  * function should only be executed once by the boot processor before the
  68  * kernel command line parameters are parsed, because otherwise it is not
  69  * possible to override this in hardlockup_panic_setup().
  70  */
  71 void watchdog_enable_hardlockup_detector(bool val)
  72 {
  73         hardlockup_detector_enabled = val;
  74 }
  75
  76 bool watchdog_hardlockup_detector_is_enabled(void)
  77 {
  78         return hardlockup_detector_enabled;
  79 }
  80
  81 static int __init hardlockup_panic_setup(char *str)
  82 {
  83         if (!strncmp(str, "panic", 5))
  84                 hardlockup_panic = 1;
  85         else if (!strncmp(str, "nopanic", 7))
  86                 hardlockup_panic = 0;
  87         else if (!strncmp(str, "0", 1))
  88                 watchdog_user_enabled = 0;
  89         else if (!strncmp(str, "1", 1) || !strncmp(str, "2", 1)) {
  90                 /*
  91                  * Setting 'nmi_watchdog=1' or 'nmi_watchdog=2' (legacy option)
  92                  * has the same effect.
  93                  */
  94                 watchdog_user_enabled = 1;
  95                 watchdog_enable_hardlockup_detector(true);
  96         }
  97         return 1;
  98 }
  99 __setup("nmi_watchdog=", hardlockup_panic_setup);
 100 #endif
 101
 102 unsigned int __read_mostly softlockup_panic =
 103                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 104
 105 static int __init softlockup_panic_setup(char *str)
 106 {
 107         softlockup_panic = simple_strtoul(str, NULL, 0);
 108
 109         return 1;
 110 }
 111 __setup("softlockup_panic=", softlockup_panic_setup);
 112
 113 static int __init nowatchdog_setup(char *str)
 114 {
 115         watchdog_user_enabled = 0;
 116         return 1;
 117 }
 118 __setup("nowatchdog", nowatchdog_setup);
 119
 120 /* deprecated */
 121 static int __init nosoftlockup_setup(char *str)
 122 {
 123         watchdog_user_enabled = 0;
 124         return 1;
 125 }
 126 __setup("nosoftlockup", nosoftlockup_setup);
 127 /*  */
 128 #ifdef CONFIG_SMP
 129 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 130 {
 131         sysctl_softlockup_all_cpu_backtrace =
 132                 !!simple_strtol(str, NULL, 0);
 133         return 1;
 134 }
 135 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 136 #endif
 137
 138 /*
 139  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 140  * lockups can have false positives under extreme conditions. So we generally
 141  * want a higher threshold for soft lockups than for hard lockups. So we couple
 142  * the thresholds with a factor: we make the soft threshold twice the amount of
 143  * time the hard threshold is.
 144  */
 145 static int get_softlockup_thresh(void)
 146 {
 147         return watchdog_thresh * 2;
 148 }
 149
 150 /*
 151  * Returns seconds, approximately.  We don't need nanosecond
 152  * resolution, and we don't need to waste time with a big divide when
 153  * 2^30ns == 1.074s.
 154  */
 155 static unsigned long get_timestamp(void)
 156 {
 157         return local_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 158 }
 159
 160 static void set_sample_period(void)
 161 {
 162         /*
 163          * convert watchdog_thresh from seconds to ns
 164          * the divide by 5 is to give hrtimer several chances (two
 165          * or three with the current relation between the soft
 166          * and hard thresholds) to increment before the
 167          * hardlockup detector generates a warning
 168          */
 169         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 170 }
 171
 172 /* Commands for resetting the watchdog */
 173 static void __touch_watchdog(void)
 174 {
 175         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 176 }
 177
 178 void touch_softlockup_watchdog(void)
 179 {
 180         /*
 181          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 182          * gets zeroed here, so use the raw_ operation.
 183          */
 184         raw_cpu_write(watchdog_touch_ts, 0);
 185 }
 186 EXPORT_SYMBOL(touch_softlockup_watchdog);
 187
 188 void touch_all_softlockup_watchdogs(void)
 189 {
 190         int cpu;
 191
 192         /*
 193          * this is done lockless
 194          * do we care if a 0 races with a timestamp?
 195          * all it means is the softlock check starts one cycle later
 196          */
 197         for_each_online_cpu(cpu)
 198                 per_cpu(watchdog_touch_ts, cpu) = 0;
 199 }
 200
 201 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 202 void touch_nmi_watchdog(void)
 203 {
 204         /*
 205          * Using __raw here because some code paths have
 206          * preemption enabled.  If preemption is enabled
 207          * then interrupts should be enabled too, in which
 208          * case we shouldn't have to worry about the watchdog
 209          * going off.
 210          */
 211         raw_cpu_write(watchdog_nmi_touch, true);
 212         touch_softlockup_watchdog();
 213 }
 214 EXPORT_SYMBOL(touch_nmi_watchdog);
 215
 216 #endif
 217
 218 void touch_softlockup_watchdog_sync(void)
 219 {
 220         __this_cpu_write(softlockup_touch_sync, true);
 221         __this_cpu_write(watchdog_touch_ts, 0);
 222 }
 223
 224 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 225 /* watchdog detector functions */
 226 static int is_hardlockup(void)
 227 {
 228         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 229
 230         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 231                 return 1;
 232
 233         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 234         return 0;
 235 }
 236 #endif
 237
 238 static int is_softlockup(unsigned long touch_ts)
 239 {
 240         unsigned long now = get_timestamp();
 241
 242         /* Warn about unreasonable delays: */
 243         if (time_after(now, touch_ts + get_softlockup_thresh()))
 244                 return now - touch_ts;
 245
 246         return 0;
 247 }
 248
 249 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 250
 251 static struct perf_event_attr wd_hw_attr = {
 252         .type           = PERF_TYPE_HARDWARE,
 253         .config         = PERF_COUNT_HW_CPU_CYCLES,
 254         .size           = sizeof(struct perf_event_attr),
 255         .pinned         = 1,
 256         .disabled       = 1,
 257 };
 258
 259 /* Callback function for perf event subsystem */
 260 static void watchdog_overflow_callback(struct perf_event *event,
 261                  struct perf_sample_data *data,
 262                  struct pt_regs *regs)
 263 {
 264         /* Ensure the watchdog never gets throttled */
 265         event->hw.interrupts = 0;
 266
 267         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 268                 __this_cpu_write(watchdog_nmi_touch, false);
 269                 return;
 270         }
 271
 272         /* check for a hardlockup
 273          * This is done by making sure our timer interrupt
 274          * is incrementing.  The timer interrupt should have
 275          * fired multiple times before we overflow'd.  If it hasn't
 276          * then this is a good indication the cpu is stuck
 277          */
 278         if (is_hardlockup()) {
 279                 int this_cpu = smp_processor_id();
 280
 281                 /* only print hardlockups once */
 282                 if (__this_cpu_read(hard_watchdog_warn) == true)
 283                         return;
 284
 285                 if (hardlockup_panic)
 286                         panic("Watchdog detected hard LOCKUP on cpu %d",
 287                               this_cpu);
 288                 else
 289                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
 290                              this_cpu);
 291
 292                 __this_cpu_write(hard_watchdog_warn, true);
 293                 return;
 294         }
 295
 296         __this_cpu_write(hard_watchdog_warn, false);
 297         return;
 298 }
 299 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 300
 301 static void watchdog_interrupt_count(void)
 302 {
 303         __this_cpu_inc(hrtimer_interrupts);
 304 }
 305
 306 static int watchdog_nmi_enable(unsigned int cpu);
 307 static void watchdog_nmi_disable(unsigned int cpu);
 308
 309 /* watchdog kicker functions */
 310 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 311 {
 312         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 313         struct pt_regs *regs = get_irq_regs();
 314         int duration;
 315         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 316
 317         /* kick the hardlockup detector */
 318         watchdog_interrupt_count();
 319
 320         /* kick the softlockup detector */
 321         wake_up_process(__this_cpu_read(softlockup_watchdog));
 322
 323         /* .. and repeat */
 324         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 325
 326         if (touch_ts == 0) {
 327                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 328                         /*
 329                          * If the time stamp was touched atomically
 330                          * make sure the scheduler tick is up to date.
 331                          */
 332                         __this_cpu_write(softlockup_touch_sync, false);
 333                         sched_clock_tick();
 334                 }
 335
 336                 /* Clear the guest paused flag on watchdog reset */
 337                 kvm_check_and_clear_guest_paused();
 338                 __touch_watchdog();
 339                 return HRTIMER_RESTART;
 340         }
 341
 342         /* check for a softlockup
 343          * This is done by making sure a high priority task is
 344          * being scheduled.  The task touches the watchdog to
 345          * indicate it is getting cpu time.  If it hasn't then
 346          * this is a good indication some task is hogging the cpu
 347          */
 348         duration = is_softlockup(touch_ts);
 349         if (unlikely(duration)) {
 350                 /*
 351                  * If a virtual machine is stopped by the host it can look to
 352                  * the watchdog like a soft lockup, check to see if the host
 353                  * stopped the vm before we issue the warning
 354                  */
 355                 if (kvm_check_and_clear_guest_paused())
 356                         return HRTIMER_RESTART;
 357
 358                 /* only warn once */
 359                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 360                         /*
 361                          * When multiple processes are causing softlockups the
 362                          * softlockup detector only warns on the first one
 363                          * because the code relies on a full quiet cycle to
 364                          * re-arm.  The second process prevents the quiet cycle
 365                          * and never gets reported.  Use task pointers to detect
 366                          * this.
 367                          */
 368                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 369                             current) {
 370                                 __this_cpu_write(soft_watchdog_warn, false);
 371                                 __touch_watchdog();
 372                         }
 373                         return HRTIMER_RESTART;
 374                 }
 375
 376                 if (softlockup_all_cpu_backtrace) {
 377                         /* Prevent multiple soft-lockup reports if one cpu is already
 378                          * engaged in dumping cpu back traces
 379                          */
 380                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 381                                 /* Someone else will report us. Let's give up */
 382                                 __this_cpu_write(soft_watchdog_warn, true);
 383                                 return HRTIMER_RESTART;
 384                         }
 385                 }
 386
 387                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 388                         smp_processor_id(), duration,
 389                         current->comm, task_pid_nr(current));
 390                 __this_cpu_write(softlockup_task_ptr_saved, current);
 391                 print_modules();
 392                 print_irqtrace_events(current);
 393                 if (regs)
 394                         show_regs(regs);
 395                 else
 396                         dump_stack();
 397
 398                 if (softlockup_all_cpu_backtrace) {
 399                         /* Avoid generating two back traces for current
 400                          * given that one is already made above
 401                          */
 402                         trigger_allbutself_cpu_backtrace();
 403
 404                         clear_bit(0, &soft_lockup_nmi_warn);
 405                         /* Barrier to sync with other cpus */
 406                         smp_mb__after_atomic();
 407                 }
 408
 409                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 410                 if (softlockup_panic)
 411                         panic("softlockup: hung tasks");
 412                 __this_cpu_write(soft_watchdog_warn, true);
 413         } else
 414                 __this_cpu_write(soft_watchdog_warn, false);
 415
 416         return HRTIMER_RESTART;
 417 }
 418
 419 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 420 {
 421         struct sched_param param = { .sched_priority = prio };
 422
 423         sched_setscheduler(current, policy, &param);
 424 }
 425
 426 static void watchdog_enable(unsigned int cpu)
 427 {
 428         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 429
 430         /* kick off the timer for the hardlockup detector */
 431         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 432         hrtimer->function = watchdog_timer_fn;
 433
 434         /* Enable the perf event */
 435         watchdog_nmi_enable(cpu);
 436
 437         /* done here because hrtimer_start can only pin to smp_processor_id() */
 438         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 439                       HRTIMER_MODE_REL_PINNED);
 440
 441         /* initialize timestamp */
 442         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 443         __touch_watchdog();
 444 }
 445
 446 static void watchdog_disable(unsigned int cpu)
 447 {
 448         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 449
 450         watchdog_set_prio(SCHED_NORMAL, 0);
 451         hrtimer_cancel(hrtimer);
 452         /* disable the perf event */
 453         watchdog_nmi_disable(cpu);
 454 }
 455
 456 static void watchdog_cleanup(unsigned int cpu, bool online)
 457 {
 458         watchdog_disable(cpu);
 459 }
 460
 461 static int watchdog_should_run(unsigned int cpu)
 462 {
 463         return __this_cpu_read(hrtimer_interrupts) !=
 464                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 465 }
 466
 467 /*
 468  * The watchdog thread function - touches the timestamp.
 469  *
 470  * It only runs once every sample_period seconds (4 seconds by
 471  * default) to reset the softlockup timestamp. If this gets delayed
 472  * for more than 2*watchdog_thresh seconds then the debug-printout
 473  * triggers in watchdog_timer_fn().
 474  */
 475 static void watchdog(unsigned int cpu)
 476 {
 477         __this_cpu_write(soft_lockup_hrtimer_cnt,
 478                          __this_cpu_read(hrtimer_interrupts));
 479         __touch_watchdog();
 480 }
 481
 482 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 483 /*
 484  * People like the simple clean cpu node info on boot.
 485  * Reduce the watchdog noise by only printing messages
 486  * that are different from what cpu0 displayed.
 487  */
 488 static unsigned long cpu0_err;
 489
 490 static int watchdog_nmi_enable(unsigned int cpu)
 491 {
 492         struct perf_event_attr *wd_attr;
 493         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 494
 495         /*
 496          * Some kernels need to default hard lockup detection to
 497          * 'disabled', for example a guest on a hypervisor.
 498          */
 499         if (!watchdog_hardlockup_detector_is_enabled()) {
 500                 event = ERR_PTR(-ENOENT);
 501                 goto handle_err;
 502         }
 503
 504         /* is it already setup and enabled? */
 505         if (event && event->state > PERF_EVENT_STATE_OFF)
 506                 goto out;
 507
 508         /* it is setup but not enabled */
 509         if (event != NULL)
 510                 goto out_enable;
 511
 512         wd_attr = &wd_hw_attr;
 513         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 514
 515         /* Try to register using hardware perf events */
 516         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 517
 518 handle_err:
 519         /* save cpu0 error for future comparision */
 520         if (cpu == 0 && IS_ERR(event))
 521                 cpu0_err = PTR_ERR(event);
 522
 523         if (!IS_ERR(event)) {
 524                 /* only print for cpu0 or different than cpu0 */
 525                 if (cpu == 0 || cpu0_err)
 526                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 527                 goto out_save;
 528         }
 529
 530         /* skip displaying the same error again */
 531         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 532                 return PTR_ERR(event);
 533
 534         /* vary the KERN level based on the returned errno */
 535         if (PTR_ERR(event) == -EOPNOTSUPP)
 536                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 537         else if (PTR_ERR(event) == -ENOENT)
 538                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 539                          cpu);
 540         else
 541                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 542                         cpu, PTR_ERR(event));
 543         return PTR_ERR(event);
 544
 545         /* success path */
 546 out_save:
 547         per_cpu(watchdog_ev, cpu) = event;
 548 out_enable:
 549         perf_event_enable(per_cpu(watchdog_ev, cpu));
 550 out:
 551         return 0;
 552 }
 553
 554 static void watchdog_nmi_disable(unsigned int cpu)
 555 {
 556         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 557
 558         if (event) {
 559                 perf_event_disable(event);
 560                 per_cpu(watchdog_ev, cpu) = NULL;
 561
 562                 /* should be in cleanup, but blocks oprofile */
 563                 perf_event_release_kernel(event);
 564         }
 565         if (cpu == 0) {
 566                 /* watchdog_nmi_enable() expects this to be zero initially. */
 567                 cpu0_err = 0;
 568         }
 569 }
 570 #else
 571 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 572 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 573 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 574
 575 static struct smp_hotplug_thread watchdog_threads = {
 576         .store                  = &softlockup_watchdog,
 577         .thread_should_run      = watchdog_should_run,
 578         .thread_fn              = watchdog,
 579         .thread_comm            = "watchdog/%u",
 580         .setup                  = watchdog_enable,
 581         .cleanup                = watchdog_cleanup,
 582         .park                   = watchdog_disable,
 583         .unpark                 = watchdog_enable,
 584 };
 585
 586 static void restart_watchdog_hrtimer(void *info)
 587 {
 588         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 589         int ret;
 590
 591         /*
 592          * No need to cancel and restart hrtimer if it is currently executing
 593          * because it will reprogram itself with the new period now.
 594          * We should never see it unqueued here because we are running per-cpu
 595          * with interrupts disabled.
 596          */
 597         ret = hrtimer_try_to_cancel(hrtimer);
 598         if (ret == 1)
 599                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 600                                 HRTIMER_MODE_REL_PINNED);
 601 }
 602
 603 static void update_timers(int cpu)
 604 {
 605         /*
 606          * Make sure that perf event counter will adopt to a new
 607          * sampling period. Updating the sampling period directly would
 608          * be much nicer but we do not have an API for that now so
 609          * let's use a big hammer.
 610          * Hrtimer will adopt the new period on the next tick but this
 611          * might be late already so we have to restart the timer as well.
 612          */
 613         watchdog_nmi_disable(cpu);
 614         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 615         watchdog_nmi_enable(cpu);
 616 }
 617
 618 static void update_timers_all_cpus(void)
 619 {
 620         int cpu;
 621
 622         get_online_cpus();
 623         for_each_online_cpu(cpu)
 624                 update_timers(cpu);
 625         put_online_cpus();
 626 }
 627
 628 static int watchdog_enable_all_cpus(bool sample_period_changed)
 629 {
 630         int err = 0;
 631
 632         if (!watchdog_running) {
 633                 err = smpboot_register_percpu_thread(&watchdog_threads);
 634                 if (err)
 635                         pr_err("Failed to create watchdog threads, disabled\n");
 636                 else
 637                         watchdog_running = 1;
 638         } else if (sample_period_changed) {
 639                 update_timers_all_cpus();
 640         }
 641
 642         return err;
 643 }
 644
 645 /* prepare/enable/disable routines */
 646 /* sysctl functions */
 647 #ifdef CONFIG_SYSCTL
 648 static void watchdog_disable_all_cpus(void)
 649 {
 650         if (watchdog_running) {
 651                 watchdog_running = 0;
 652                 smpboot_unregister_percpu_thread(&watchdog_threads);
 653         }
 654 }
 655
 656 /*
 657  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 658  */
 659
 660 int proc_dowatchdog(struct ctl_table *table, int write,
 661                     void __user *buffer, size_t *lenp, loff_t *ppos)
 662 {
 663         int err, old_thresh, old_enabled;
 664         bool old_hardlockup;
 665         static DEFINE_MUTEX(watchdog_proc_mutex);
 666
 667         mutex_lock(&watchdog_proc_mutex);
 668         old_thresh = ACCESS_ONCE(watchdog_thresh);
 669         old_enabled = ACCESS_ONCE(watchdog_user_enabled);
 670         old_hardlockup = watchdog_hardlockup_detector_is_enabled();
 671
 672         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 673         if (err || !write)
 674                 goto out;
 675
 676         set_sample_period();
 677         /*
 678          * Watchdog threads shouldn't be enabled if they are
 679          * disabled. The 'watchdog_running' variable check in
 680          * watchdog_*_all_cpus() function takes care of this.
 681          */
 682         if (watchdog_user_enabled && watchdog_thresh) {
 683                 /*
 684                  * Prevent a change in watchdog_thresh accidentally overriding
 685                  * the enablement of the hardlockup detector.
 686                  */
 687                 if (watchdog_user_enabled != old_enabled)
 688                         watchdog_enable_hardlockup_detector(true);
 689                 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
 690         } else
 691                 watchdog_disable_all_cpus();
 692
 693         /* Restore old values on failure */
 694         if (err) {
 695                 watchdog_thresh = old_thresh;
 696                 watchdog_user_enabled = old_enabled;
 697                 watchdog_enable_hardlockup_detector(old_hardlockup);
 698         }
 699 out:
 700         mutex_unlock(&watchdog_proc_mutex);
 701         return err;
 702 }
 703 #endif /* CONFIG_SYSCTL */
 704
 705 void __init lockup_detector_init(void)
 706 {
 707         set_sample_period();
 708
 709         if (watchdog_user_enabled)
 710                 watchdog_enable_all_cpus(false);
 711 }