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) "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 #include <uapi/linux/sched/types.h>
  23 #include <linux/tick.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/sched/clock.h>
  26 #include <linux/sched/debug.h>
  27
  28 #include <asm/irq_regs.h>
  29 #include <linux/kvm_para.h>
  30 #include <linux/kthread.h>
  31
  32 static DEFINE_MUTEX(watchdog_mutex);
  33
  34 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
  35 # define WATCHDOG_DEFAULT       (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
  36 # define NMI_WATCHDOG_DEFAULT   1
  37 #else
  38 # define WATCHDOG_DEFAULT       (SOFT_WATCHDOG_ENABLED)
  39 # define NMI_WATCHDOG_DEFAULT   0
  40 #endif
  41
  42 unsigned long __read_mostly watchdog_enabled;
  43 int __read_mostly watchdog_user_enabled = 1;
  44 int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
  45 int __read_mostly soft_watchdog_user_enabled = 1;
  46 int __read_mostly watchdog_thresh = 10;
  47 int __read_mostly nmi_watchdog_available;
  48
  49 struct cpumask watchdog_allowed_mask __read_mostly;
  50 static bool softlockup_threads_initialized __read_mostly;
  51
  52 struct cpumask watchdog_cpumask __read_mostly;
  53 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  54
  55 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  56 /*
  57  * Should we panic when a soft-lockup or hard-lockup occurs:
  58  */
  59 unsigned int __read_mostly hardlockup_panic =
  60                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  61 /*
  62  * We may not want to enable hard lockup detection by default in all cases,
  63  * for example when running the kernel as a guest on a hypervisor. In these
  64  * cases this function can be called to disable hard lockup detection. This
  65  * function should only be executed once by the boot processor before the
  66  * kernel command line parameters are parsed, because otherwise it is not
  67  * possible to override this in hardlockup_panic_setup().
  68  */
  69 void __init hardlockup_detector_disable(void)
  70 {
  71         nmi_watchdog_user_enabled = 0;
  72 }
  73
  74 static int __init hardlockup_panic_setup(char *str)
  75 {
  76         if (!strncmp(str, "panic", 5))
  77                 hardlockup_panic = 1;
  78         else if (!strncmp(str, "nopanic", 7))
  79                 hardlockup_panic = 0;
  80         else if (!strncmp(str, "0", 1))
  81                 nmi_watchdog_user_enabled = 0;
  82         else if (!strncmp(str, "1", 1))
  83                 nmi_watchdog_user_enabled = 1;
  84         return 1;
  85 }
  86 __setup("nmi_watchdog=", hardlockup_panic_setup);
  87
  88 # ifdef CONFIG_SMP
  89 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  90
  91 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
  92 {
  93         sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
  94         return 1;
  95 }
  96 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
  97 # endif /* CONFIG_SMP */
  98 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
  99
 100 /*
 101  * These functions can be overridden if an architecture implements its
 102  * own hardlockup detector.
 103  *
 104  * watchdog_nmi_enable/disable can be implemented to start and stop when
 105  * softlockup watchdog threads start and stop. The arch must select the
 106  * SOFTLOCKUP_DETECTOR Kconfig.
 107  */
 108 int __weak watchdog_nmi_enable(unsigned int cpu)
 109 {
 110         hardlockup_detector_perf_enable();
 111         return 0;
 112 }
 113
 114 void __weak watchdog_nmi_disable(unsigned int cpu)
 115 {
 116         hardlockup_detector_perf_disable();
 117 }
 118
 119 /* Return 0, if a NMI watchdog is available. Error code otherwise */
 120 int __weak __init watchdog_nmi_probe(void)
 121 {
 122         return hardlockup_detector_perf_init();
 123 }
 124
 125 /**
 126  * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs
 127  * @run:        If false stop the watchdogs on all enabled CPUs
 128  *              If true start the watchdogs on all enabled CPUs
 129  *
 130  * The core call order is:
 131  * watchdog_nmi_reconfigure(false);
 132  * update_variables();
 133  * watchdog_nmi_reconfigure(true);
 134  *
 135  * The second call which starts the watchdogs again guarantees that the
 136  * following variables are stable across the call.
 137  * - watchdog_enabled
 138  * - watchdog_thresh
 139  * - watchdog_cpumask
 140  *
 141  * After the call the variables can be changed again.
 142  */
 143 void __weak watchdog_nmi_reconfigure(bool run) { }
 144
 145 /**
 146  * lockup_detector_update_enable - Update the sysctl enable bit
 147  *
 148  * Caller needs to make sure that the NMI/perf watchdogs are off, so this
 149  * can't race with watchdog_nmi_disable().
 150  */
 151 static void lockup_detector_update_enable(void)
 152 {
 153         watchdog_enabled = 0;
 154         if (!watchdog_user_enabled)
 155                 return;
 156         if (nmi_watchdog_available && nmi_watchdog_user_enabled)
 157                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 158         if (soft_watchdog_user_enabled)
 159                 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
 160 }
 161
 162 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 163
 164 /* Global variables, exported for sysctl */
 165 unsigned int __read_mostly softlockup_panic =
 166                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 167
 168 static u64 __read_mostly sample_period;
 169
 170 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 171 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
 172 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 173 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 174 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 175 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 176 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 177 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 178 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 179 static unsigned long soft_lockup_nmi_warn;
 180
 181 static int __init softlockup_panic_setup(char *str)
 182 {
 183         softlockup_panic = simple_strtoul(str, NULL, 0);
 184         return 1;
 185 }
 186 __setup("softlockup_panic=", softlockup_panic_setup);
 187
 188 static int __init nowatchdog_setup(char *str)
 189 {
 190         watchdog_user_enabled = 0;
 191         return 1;
 192 }
 193 __setup("nowatchdog", nowatchdog_setup);
 194
 195 static int __init nosoftlockup_setup(char *str)
 196 {
 197         soft_watchdog_user_enabled = 0;
 198         return 1;
 199 }
 200 __setup("nosoftlockup", nosoftlockup_setup);
 201
 202 #ifdef CONFIG_SMP
 203 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 204
 205 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 206 {
 207         sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
 208         return 1;
 209 }
 210 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 211 #endif
 212
 213 static void __lockup_detector_cleanup(void);
 214
 215 /*
 216  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 217  * lockups can have false positives under extreme conditions. So we generally
 218  * want a higher threshold for soft lockups than for hard lockups. So we couple
 219  * the thresholds with a factor: we make the soft threshold twice the amount of
 220  * time the hard threshold is.
 221  */
 222 static int get_softlockup_thresh(void)
 223 {
 224         return watchdog_thresh * 2;
 225 }
 226
 227 /*
 228  * Returns seconds, approximately.  We don't need nanosecond
 229  * resolution, and we don't need to waste time with a big divide when
 230  * 2^30ns == 1.074s.
 231  */
 232 static unsigned long get_timestamp(void)
 233 {
 234         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 235 }
 236
 237 static void set_sample_period(void)
 238 {
 239         /*
 240          * convert watchdog_thresh from seconds to ns
 241          * the divide by 5 is to give hrtimer several chances (two
 242          * or three with the current relation between the soft
 243          * and hard thresholds) to increment before the
 244          * hardlockup detector generates a warning
 245          */
 246         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 247         watchdog_update_hrtimer_threshold(sample_period);
 248 }
 249
 250 /* Commands for resetting the watchdog */
 251 static void __touch_watchdog(void)
 252 {
 253         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 254 }
 255
 256 /**
 257  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 258  *
 259  * Call when the scheduler may have stalled for legitimate reasons
 260  * preventing the watchdog task from executing - e.g. the scheduler
 261  * entering idle state.  This should only be used for scheduler events.
 262  * Use touch_softlockup_watchdog() for everything else.
 263  */
 264 void touch_softlockup_watchdog_sched(void)
 265 {
 266         /*
 267          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 268          * gets zeroed here, so use the raw_ operation.
 269          */
 270         raw_cpu_write(watchdog_touch_ts, 0);
 271 }
 272
 273 void touch_softlockup_watchdog(void)
 274 {
 275         touch_softlockup_watchdog_sched();
 276         wq_watchdog_touch(raw_smp_processor_id());
 277 }
 278 EXPORT_SYMBOL(touch_softlockup_watchdog);
 279
 280 void touch_all_softlockup_watchdogs(void)
 281 {
 282         int cpu;
 283
 284         /*
 285          * watchdog_mutex cannpt be taken here, as this might be called
 286          * from (soft)interrupt context, so the access to
 287          * watchdog_allowed_cpumask might race with a concurrent update.
 288          *
 289          * The watchdog time stamp can race against a concurrent real
 290          * update as well, the only side effect might be a cycle delay for
 291          * the softlockup check.
 292          */
 293         for_each_cpu(cpu, &watchdog_allowed_mask)
 294                 per_cpu(watchdog_touch_ts, cpu) = 0;
 295         wq_watchdog_touch(-1);
 296 }
 297
 298 void touch_softlockup_watchdog_sync(void)
 299 {
 300         __this_cpu_write(softlockup_touch_sync, true);
 301         __this_cpu_write(watchdog_touch_ts, 0);
 302 }
 303
 304 static int is_softlockup(unsigned long touch_ts)
 305 {
 306         unsigned long now = get_timestamp();
 307
 308         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 309                 /* Warn about unreasonable delays. */
 310                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 311                         return now - touch_ts;
 312         }
 313         return 0;
 314 }
 315
 316 /* watchdog detector functions */
 317 bool is_hardlockup(void)
 318 {
 319         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 320
 321         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 322                 return true;
 323
 324         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 325         return false;
 326 }
 327
 328 static void watchdog_interrupt_count(void)
 329 {
 330         __this_cpu_inc(hrtimer_interrupts);
 331 }
 332
 333 /* watchdog kicker functions */
 334 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 335 {
 336         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 337         struct pt_regs *regs = get_irq_regs();
 338         int duration;
 339         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 340
 341         if (!watchdog_enabled)
 342                 return HRTIMER_NORESTART;
 343
 344         /* kick the hardlockup detector */
 345         watchdog_interrupt_count();
 346
 347         /* kick the softlockup detector */
 348         wake_up_process(__this_cpu_read(softlockup_watchdog));
 349
 350         /* .. and repeat */
 351         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 352
 353         if (touch_ts == 0) {
 354                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 355                         /*
 356                          * If the time stamp was touched atomically
 357                          * make sure the scheduler tick is up to date.
 358                          */
 359                         __this_cpu_write(softlockup_touch_sync, false);
 360                         sched_clock_tick();
 361                 }
 362
 363                 /* Clear the guest paused flag on watchdog reset */
 364                 kvm_check_and_clear_guest_paused();
 365                 __touch_watchdog();
 366                 return HRTIMER_RESTART;
 367         }
 368
 369         /* check for a softlockup
 370          * This is done by making sure a high priority task is
 371          * being scheduled.  The task touches the watchdog to
 372          * indicate it is getting cpu time.  If it hasn't then
 373          * this is a good indication some task is hogging the cpu
 374          */
 375         duration = is_softlockup(touch_ts);
 376         if (unlikely(duration)) {
 377                 /*
 378                  * If a virtual machine is stopped by the host it can look to
 379                  * the watchdog like a soft lockup, check to see if the host
 380                  * stopped the vm before we issue the warning
 381                  */
 382                 if (kvm_check_and_clear_guest_paused())
 383                         return HRTIMER_RESTART;
 384
 385                 /* only warn once */
 386                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 387                         /*
 388                          * When multiple processes are causing softlockups the
 389                          * softlockup detector only warns on the first one
 390                          * because the code relies on a full quiet cycle to
 391                          * re-arm.  The second process prevents the quiet cycle
 392                          * and never gets reported.  Use task pointers to detect
 393                          * this.
 394                          */
 395                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 396                             current) {
 397                                 __this_cpu_write(soft_watchdog_warn, false);
 398                                 __touch_watchdog();
 399                         }
 400                         return HRTIMER_RESTART;
 401                 }
 402
 403                 if (softlockup_all_cpu_backtrace) {
 404                         /* Prevent multiple soft-lockup reports if one cpu is already
 405                          * engaged in dumping cpu back traces
 406                          */
 407                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 408                                 /* Someone else will report us. Let's give up */
 409                                 __this_cpu_write(soft_watchdog_warn, true);
 410                                 return HRTIMER_RESTART;
 411                         }
 412                 }
 413
 414                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 415                         smp_processor_id(), duration,
 416                         current->comm, task_pid_nr(current));
 417                 __this_cpu_write(softlockup_task_ptr_saved, current);
 418                 print_modules();
 419                 print_irqtrace_events(current);
 420                 if (regs)
 421                         show_regs(regs);
 422                 else
 423                         dump_stack();
 424
 425                 if (softlockup_all_cpu_backtrace) {
 426                         /* Avoid generating two back traces for current
 427                          * given that one is already made above
 428                          */
 429                         trigger_allbutself_cpu_backtrace();
 430
 431                         clear_bit(0, &soft_lockup_nmi_warn);
 432                         /* Barrier to sync with other cpus */
 433                         smp_mb__after_atomic();
 434                 }
 435
 436                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 437                 if (softlockup_panic)
 438                         panic("softlockup: hung tasks");
 439                 __this_cpu_write(soft_watchdog_warn, true);
 440         } else
 441                 __this_cpu_write(soft_watchdog_warn, false);
 442
 443         return HRTIMER_RESTART;
 444 }
 445
 446 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 447 {
 448         struct sched_param param = { .sched_priority = prio };
 449
 450         sched_setscheduler(current, policy, &param);
 451 }
 452
 453 static void watchdog_enable(unsigned int cpu)
 454 {
 455         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 456
 457         /*
 458          * Start the timer first to prevent the NMI watchdog triggering
 459          * before the timer has a chance to fire.
 460          */
 461         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 462         hrtimer->function = watchdog_timer_fn;
 463         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 464                       HRTIMER_MODE_REL_PINNED);
 465
 466         /* Initialize timestamp */
 467         __touch_watchdog();
 468         /* Enable the perf event */
 469         if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
 470                 watchdog_nmi_enable(cpu);
 471
 472         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 473 }
 474
 475 static void watchdog_disable(unsigned int cpu)
 476 {
 477         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 478
 479         watchdog_set_prio(SCHED_NORMAL, 0);
 480         /*
 481          * Disable the perf event first. That prevents that a large delay
 482          * between disabling the timer and disabling the perf event causes
 483          * the perf NMI to detect a false positive.
 484          */
 485         watchdog_nmi_disable(cpu);
 486         hrtimer_cancel(hrtimer);
 487 }
 488
 489 static void watchdog_cleanup(unsigned int cpu, bool online)
 490 {
 491         watchdog_disable(cpu);
 492 }
 493
 494 static int watchdog_should_run(unsigned int cpu)
 495 {
 496         return __this_cpu_read(hrtimer_interrupts) !=
 497                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 498 }
 499
 500 /*
 501  * The watchdog thread function - touches the timestamp.
 502  *
 503  * It only runs once every sample_period seconds (4 seconds by
 504  * default) to reset the softlockup timestamp. If this gets delayed
 505  * for more than 2*watchdog_thresh seconds then the debug-printout
 506  * triggers in watchdog_timer_fn().
 507  */
 508 static void watchdog(unsigned int cpu)
 509 {
 510         __this_cpu_write(soft_lockup_hrtimer_cnt,
 511                          __this_cpu_read(hrtimer_interrupts));
 512         __touch_watchdog();
 513 }
 514
 515 static struct smp_hotplug_thread watchdog_threads = {
 516         .store                  = &softlockup_watchdog,
 517         .thread_should_run      = watchdog_should_run,
 518         .thread_fn              = watchdog,
 519         .thread_comm            = "watchdog/%u",
 520         .setup                  = watchdog_enable,
 521         .cleanup                = watchdog_cleanup,
 522         .park                   = watchdog_disable,
 523         .unpark                 = watchdog_enable,
 524 };
 525
 526 static void softlockup_update_smpboot_threads(void)
 527 {
 528         lockdep_assert_held(&watchdog_mutex);
 529
 530         if (!softlockup_threads_initialized)
 531                 return;
 532
 533         smpboot_update_cpumask_percpu_thread(&watchdog_threads,
 534                                              &watchdog_allowed_mask);
 535         __lockup_detector_cleanup();
 536 }
 537
 538 /* Temporarily park all watchdog threads */
 539 static void softlockup_park_all_threads(void)
 540 {
 541         cpumask_clear(&watchdog_allowed_mask);
 542         softlockup_update_smpboot_threads();
 543 }
 544
 545 /* Unpark enabled threads */
 546 static void softlockup_unpark_threads(void)
 547 {
 548         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 549         softlockup_update_smpboot_threads();
 550 }
 551
 552 static void softlockup_reconfigure_threads(void)
 553 {
 554         watchdog_nmi_reconfigure(false);
 555         softlockup_park_all_threads();
 556         set_sample_period();
 557         lockup_detector_update_enable();
 558         if (watchdog_enabled && watchdog_thresh)
 559                 softlockup_unpark_threads();
 560         watchdog_nmi_reconfigure(true);
 561 }
 562
 563 /*
 564  * Create the watchdog thread infrastructure.
 565  *
 566  * The threads are not unparked as watchdog_allowed_mask is empty.  When
 567  * the threads are sucessfully initialized, take the proper locks and
 568  * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
 569  */
 570 static __init void softlockup_init_threads(void)
 571 {
 572         int ret;
 573
 574         /*
 575          * If sysctl is off and watchdog got disabled on the command line,
 576          * nothing to do here.
 577          */
 578         lockup_detector_update_enable();
 579
 580         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 581             !(watchdog_enabled && watchdog_thresh))
 582                 return;
 583
 584         ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 585                                                      &watchdog_allowed_mask);
 586         if (ret) {
 587                 pr_err("Failed to initialize soft lockup detector threads\n");
 588                 return;
 589         }
 590
 591         mutex_lock(&watchdog_mutex);
 592         softlockup_threads_initialized = true;
 593         softlockup_reconfigure_threads();
 594         mutex_unlock(&watchdog_mutex);
 595 }
 596
 597 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 598 static inline int watchdog_park_threads(void) { return 0; }
 599 static inline void watchdog_unpark_threads(void) { }
 600 static inline int watchdog_enable_all_cpus(void) { return 0; }
 601 static inline void watchdog_disable_all_cpus(void) { }
 602 static inline void softlockup_init_threads(void) { }
 603 static void softlockup_reconfigure_threads(void)
 604 {
 605         watchdog_nmi_reconfigure(false);
 606         lockup_detector_update_enable();
 607         watchdog_nmi_reconfigure(true);
 608 }
 609 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 610
 611 static void __lockup_detector_cleanup(void)
 612 {
 613         lockdep_assert_held(&watchdog_mutex);
 614         hardlockup_detector_perf_cleanup();
 615 }
 616
 617 /**
 618  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 619  *
 620  * Caller must not hold the cpu hotplug rwsem.
 621  */
 622 void lockup_detector_cleanup(void)
 623 {
 624         mutex_lock(&watchdog_mutex);
 625         __lockup_detector_cleanup();
 626         mutex_unlock(&watchdog_mutex);
 627 }
 628
 629 /**
 630  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 631  *
 632  * Special interface for parisc. It prevents lockup detector warnings from
 633  * the default pm_poweroff() function which busy loops forever.
 634  */
 635 void lockup_detector_soft_poweroff(void)
 636 {
 637         watchdog_enabled = 0;
 638 }
 639
 640 #ifdef CONFIG_SYSCTL
 641
 642 /* Propagate any changes to the watchdog threads */
 643 static void proc_watchdog_update(void)
 644 {
 645         /* Remove impossible cpus to keep sysctl output clean. */
 646         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 647         softlockup_reconfigure_threads();
 648 }
 649
 650 /*
 651  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 652  *
 653  * caller             | table->data points to      | 'which'
 654  * -------------------|----------------------------|--------------------------
 655  * proc_watchdog      | watchdog_user_enabled      | NMI_WATCHDOG_ENABLED |
 656  *                    |                            | SOFT_WATCHDOG_ENABLED
 657  * -------------------|----------------------------|--------------------------
 658  * proc_nmi_watchdog  | nmi_watchdog_user_enabled  | NMI_WATCHDOG_ENABLED
 659  * -------------------|----------------------------|--------------------------
 660  * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
 661  */
 662 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 663                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 664 {
 665         int err, old, *param = table->data;
 666
 667         cpu_hotplug_disable();
 668         mutex_lock(&watchdog_mutex);
 669
 670         if (!write) {
 671                 /*
 672                  * On read synchronize the userspace interface. This is a
 673                  * racy snapshot.
 674                  */
 675                 *param = (watchdog_enabled & which) != 0;
 676                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 677         } else {
 678                 old = READ_ONCE(*param);
 679                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 680                 if (!err && old != READ_ONCE(*param))
 681                         proc_watchdog_update();
 682         }
 683         mutex_unlock(&watchdog_mutex);
 684         cpu_hotplug_enable();
 685         return err;
 686 }
 687
 688 /*
 689  * /proc/sys/kernel/watchdog
 690  */
 691 int proc_watchdog(struct ctl_table *table, int write,
 692                   void __user *buffer, size_t *lenp, loff_t *ppos)
 693 {
 694         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 695                                     table, write, buffer, lenp, ppos);
 696 }
 697
 698 /*
 699  * /proc/sys/kernel/nmi_watchdog
 700  */
 701 int proc_nmi_watchdog(struct ctl_table *table, int write,
 702                       void __user *buffer, size_t *lenp, loff_t *ppos)
 703 {
 704         if (!nmi_watchdog_available && write)
 705                 return -ENOTSUPP;
 706         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 707                                     table, write, buffer, lenp, ppos);
 708 }
 709
 710 /*
 711  * /proc/sys/kernel/soft_watchdog
 712  */
 713 int proc_soft_watchdog(struct ctl_table *table, int write,
 714                         void __user *buffer, size_t *lenp, loff_t *ppos)
 715 {
 716         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 717                                     table, write, buffer, lenp, ppos);
 718 }
 719
 720 /*
 721  * /proc/sys/kernel/watchdog_thresh
 722  */
 723 int proc_watchdog_thresh(struct ctl_table *table, int write,
 724                          void __user *buffer, size_t *lenp, loff_t *ppos)
 725 {
 726         int err, old;
 727
 728         cpu_hotplug_disable();
 729         mutex_lock(&watchdog_mutex);
 730
 731         old = READ_ONCE(watchdog_thresh);
 732         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 733
 734         if (!err && write && old != READ_ONCE(watchdog_thresh))
 735                 proc_watchdog_update();
 736
 737         mutex_unlock(&watchdog_mutex);
 738         cpu_hotplug_enable();
 739         return err;
 740 }
 741
 742 /*
 743  * The cpumask is the mask of possible cpus that the watchdog can run
 744  * on, not the mask of cpus it is actually running on.  This allows the
 745  * user to specify a mask that will include cpus that have not yet
 746  * been brought online, if desired.
 747  */
 748 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 749                           void __user *buffer, size_t *lenp, loff_t *ppos)
 750 {
 751         int err;
 752
 753         cpu_hotplug_disable();
 754         mutex_lock(&watchdog_mutex);
 755
 756         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 757         if (!err && write)
 758                 proc_watchdog_update();
 759
 760         mutex_unlock(&watchdog_mutex);
 761         cpu_hotplug_enable();
 762         return err;
 763 }
 764 #endif /* CONFIG_SYSCTL */
 765
 766 void __init lockup_detector_init(void)
 767 {
 768 #ifdef CONFIG_NO_HZ_FULL
 769         if (tick_nohz_full_enabled()) {
 770                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 771                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
 772         } else
 773                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 774 #else
 775         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 776 #endif
 777
 778         if (!watchdog_nmi_probe())
 779                 nmi_watchdog_available = true;
 780         softlockup_init_threads();
 781 }