kernel/watchdog.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Detect hard and soft lockups on a system
   4  *
   5  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   6  *
   7  * Note: Most of this code is borrowed heavily from the original softlockup
   8  * detector, so thanks to Ingo for the initial implementation.
   9  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
  10  * to those contributors as well.
  11  */
  12
  13 #define pr_fmt(fmt) "watchdog: " fmt
  14
  15 #include <linux/mm.h>
  16 #include <linux/cpu.h>
  17 #include <linux/nmi.h>
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/sysctl.h>
  21 #include <linux/tick.h>
  22 #include <linux/sched/clock.h>
  23 #include <linux/sched/debug.h>
  24 #include <linux/sched/isolation.h>
  25 #include <linux/stop_machine.h>
  26
  27 #include <asm/irq_regs.h>
  28 #include <linux/kvm_para.h>
  29
  30 static DEFINE_MUTEX(watchdog_mutex);
  31
  32 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
  33 # define WATCHDOG_DEFAULT       (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
  34 # define NMI_WATCHDOG_DEFAULT   1
  35 #else
  36 # define WATCHDOG_DEFAULT       (SOFT_WATCHDOG_ENABLED)
  37 # define NMI_WATCHDOG_DEFAULT   0
  38 #endif
  39
  40 unsigned long __read_mostly watchdog_enabled;
  41 int __read_mostly watchdog_user_enabled = 1;
  42 int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
  43 int __read_mostly soft_watchdog_user_enabled = 1;
  44 int __read_mostly watchdog_thresh = 10;
  45 static int __read_mostly nmi_watchdog_available;
  46
  47 struct cpumask watchdog_cpumask __read_mostly;
  48 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  49
  50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  51
  52 # ifdef CONFIG_SMP
  53 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  54 # endif /* CONFIG_SMP */
  55
  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 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
  89
  90 /*
  91  * These functions can be overridden if an architecture implements its
  92  * own hardlockup detector.
  93  *
  94  * watchdog_nmi_enable/disable can be implemented to start and stop when
  95  * softlockup watchdog threads start and stop. The arch must select the
  96  * SOFTLOCKUP_DETECTOR Kconfig.
  97  */
  98 int __weak watchdog_nmi_enable(unsigned int cpu)
  99 {
 100         hardlockup_detector_perf_enable();
 101         return 0;
 102 }
 103
 104 void __weak watchdog_nmi_disable(unsigned int cpu)
 105 {
 106         hardlockup_detector_perf_disable();
 107 }
 108
 109 /* Return 0, if a NMI watchdog is available. Error code otherwise */
 110 int __weak __init watchdog_nmi_probe(void)
 111 {
 112         return hardlockup_detector_perf_init();
 113 }
 114
 115 /**
 116  * watchdog_nmi_stop - Stop the watchdog for reconfiguration
 117  *
 118  * The reconfiguration steps are:
 119  * watchdog_nmi_stop();
 120  * update_variables();
 121  * watchdog_nmi_start();
 122  */
 123 void __weak watchdog_nmi_stop(void) { }
 124
 125 /**
 126  * watchdog_nmi_start - Start the watchdog after reconfiguration
 127  *
 128  * Counterpart to watchdog_nmi_stop().
 129  *
 130  * The following variables have been updated in update_variables() and
 131  * contain the currently valid configuration:
 132  * - watchdog_enabled
 133  * - watchdog_thresh
 134  * - watchdog_cpumask
 135  */
 136 void __weak watchdog_nmi_start(void) { }
 137
 138 /**
 139  * lockup_detector_update_enable - Update the sysctl enable bit
 140  *
 141  * Caller needs to make sure that the NMI/perf watchdogs are off, so this
 142  * can't race with watchdog_nmi_disable().
 143  */
 144 static void lockup_detector_update_enable(void)
 145 {
 146         watchdog_enabled = 0;
 147         if (!watchdog_user_enabled)
 148                 return;
 149         if (nmi_watchdog_available && nmi_watchdog_user_enabled)
 150                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 151         if (soft_watchdog_user_enabled)
 152                 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
 153 }
 154
 155 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 156
 157 #define SOFTLOCKUP_RESET        ULONG_MAX
 158
 159 #ifdef CONFIG_SMP
 160 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 161 #endif
 162
 163 static struct cpumask watchdog_allowed_mask __read_mostly;
 164
 165 /* Global variables, exported for sysctl */
 166 unsigned int __read_mostly softlockup_panic =
 167                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 168
 169 static bool softlockup_initialized __read_mostly;
 170 static u64 __read_mostly sample_period;
 171
 172 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 173 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 174 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 175 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 176 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 178 static unsigned long soft_lockup_nmi_warn;
 179
 180 static int __init nowatchdog_setup(char *str)
 181 {
 182         watchdog_user_enabled = 0;
 183         return 1;
 184 }
 185 __setup("nowatchdog", nowatchdog_setup);
 186
 187 static int __init nosoftlockup_setup(char *str)
 188 {
 189         soft_watchdog_user_enabled = 0;
 190         return 1;
 191 }
 192 __setup("nosoftlockup", nosoftlockup_setup);
 193
 194 static int __init watchdog_thresh_setup(char *str)
 195 {
 196         get_option(&str, &watchdog_thresh);
 197         return 1;
 198 }
 199 __setup("watchdog_thresh=", watchdog_thresh_setup);
 200
 201 static void __lockup_detector_cleanup(void);
 202
 203 /*
 204  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 205  * lockups can have false positives under extreme conditions. So we generally
 206  * want a higher threshold for soft lockups than for hard lockups. So we couple
 207  * the thresholds with a factor: we make the soft threshold twice the amount of
 208  * time the hard threshold is.
 209  */
 210 static int get_softlockup_thresh(void)
 211 {
 212         return watchdog_thresh * 2;
 213 }
 214
 215 /*
 216  * Returns seconds, approximately.  We don't need nanosecond
 217  * resolution, and we don't need to waste time with a big divide when
 218  * 2^30ns == 1.074s.
 219  */
 220 static unsigned long get_timestamp(void)
 221 {
 222         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 223 }
 224
 225 static void set_sample_period(void)
 226 {
 227         /*
 228          * convert watchdog_thresh from seconds to ns
 229          * the divide by 5 is to give hrtimer several chances (two
 230          * or three with the current relation between the soft
 231          * and hard thresholds) to increment before the
 232          * hardlockup detector generates a warning
 233          */
 234         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 235         watchdog_update_hrtimer_threshold(sample_period);
 236 }
 237
 238 /* Commands for resetting the watchdog */
 239 static void __touch_watchdog(void)
 240 {
 241         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 242 }
 243
 244 /**
 245  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 246  *
 247  * Call when the scheduler may have stalled for legitimate reasons
 248  * preventing the watchdog task from executing - e.g. the scheduler
 249  * entering idle state.  This should only be used for scheduler events.
 250  * Use touch_softlockup_watchdog() for everything else.
 251  */
 252 notrace void touch_softlockup_watchdog_sched(void)
 253 {
 254         /*
 255          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 256          * gets zeroed here, so use the raw_ operation.
 257          */
 258         raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 259 }
 260
 261 notrace void touch_softlockup_watchdog(void)
 262 {
 263         touch_softlockup_watchdog_sched();
 264         wq_watchdog_touch(raw_smp_processor_id());
 265 }
 266 EXPORT_SYMBOL(touch_softlockup_watchdog);
 267
 268 void touch_all_softlockup_watchdogs(void)
 269 {
 270         int cpu;
 271
 272         /*
 273          * watchdog_mutex cannpt be taken here, as this might be called
 274          * from (soft)interrupt context, so the access to
 275          * watchdog_allowed_cpumask might race with a concurrent update.
 276          *
 277          * The watchdog time stamp can race against a concurrent real
 278          * update as well, the only side effect might be a cycle delay for
 279          * the softlockup check.
 280          */
 281         for_each_cpu(cpu, &watchdog_allowed_mask)
 282                 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
 283         wq_watchdog_touch(-1);
 284 }
 285
 286 void touch_softlockup_watchdog_sync(void)
 287 {
 288         __this_cpu_write(softlockup_touch_sync, true);
 289         __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 290 }
 291
 292 static int is_softlockup(unsigned long touch_ts)
 293 {
 294         unsigned long now = get_timestamp();
 295
 296         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 297                 /* Warn about unreasonable delays. */
 298                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 299                         return now - touch_ts;
 300         }
 301         return 0;
 302 }
 303
 304 /* watchdog detector functions */
 305 bool is_hardlockup(void)
 306 {
 307         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 308
 309         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 310                 return true;
 311
 312         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 313         return false;
 314 }
 315
 316 static void watchdog_interrupt_count(void)
 317 {
 318         __this_cpu_inc(hrtimer_interrupts);
 319 }
 320
 321 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 322 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 323
 324 /*
 325  * The watchdog thread function - touches the timestamp.
 326  *
 327  * It only runs once every sample_period seconds (4 seconds by
 328  * default) to reset the softlockup timestamp. If this gets delayed
 329  * for more than 2*watchdog_thresh seconds then the debug-printout
 330  * triggers in watchdog_timer_fn().
 331  */
 332 static int softlockup_fn(void *data)
 333 {
 334         __touch_watchdog();
 335         complete(this_cpu_ptr(&softlockup_completion));
 336
 337         return 0;
 338 }
 339
 340 /* watchdog kicker functions */
 341 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 342 {
 343         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 344         struct pt_regs *regs = get_irq_regs();
 345         int duration;
 346         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 347
 348         if (!watchdog_enabled)
 349                 return HRTIMER_NORESTART;
 350
 351         /* kick the hardlockup detector */
 352         watchdog_interrupt_count();
 353
 354         /* kick the softlockup detector */
 355         if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 356                 reinit_completion(this_cpu_ptr(&softlockup_completion));
 357                 stop_one_cpu_nowait(smp_processor_id(),
 358                                 softlockup_fn, NULL,
 359                                 this_cpu_ptr(&softlockup_stop_work));
 360         }
 361
 362         /* .. and repeat */
 363         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 364
 365         if (touch_ts == SOFTLOCKUP_RESET) {
 366                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 367                         /*
 368                          * If the time stamp was touched atomically
 369                          * make sure the scheduler tick is up to date.
 370                          */
 371                         __this_cpu_write(softlockup_touch_sync, false);
 372                         sched_clock_tick();
 373                 }
 374
 375                 /* Clear the guest paused flag on watchdog reset */
 376                 kvm_check_and_clear_guest_paused();
 377                 __touch_watchdog();
 378                 return HRTIMER_RESTART;
 379         }
 380
 381         /* check for a softlockup
 382          * This is done by making sure a high priority task is
 383          * being scheduled.  The task touches the watchdog to
 384          * indicate it is getting cpu time.  If it hasn't then
 385          * this is a good indication some task is hogging the cpu
 386          */
 387         duration = is_softlockup(touch_ts);
 388         if (unlikely(duration)) {
 389                 /*
 390                  * If a virtual machine is stopped by the host it can look to
 391                  * the watchdog like a soft lockup, check to see if the host
 392                  * stopped the vm before we issue the warning
 393                  */
 394                 if (kvm_check_and_clear_guest_paused())
 395                         return HRTIMER_RESTART;
 396
 397                 /* only warn once */
 398                 if (__this_cpu_read(soft_watchdog_warn) == true)
 399                         return HRTIMER_RESTART;
 400
 401                 if (softlockup_all_cpu_backtrace) {
 402                         /* Prevent multiple soft-lockup reports if one cpu is already
 403                          * engaged in dumping cpu back traces
 404                          */
 405                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 406                                 /* Someone else will report us. Let's give up */
 407                                 __this_cpu_write(soft_watchdog_warn, true);
 408                                 return HRTIMER_RESTART;
 409                         }
 410                 }
 411
 412                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 413                         smp_processor_id(), duration,
 414                         current->comm, task_pid_nr(current));
 415                 print_modules();
 416                 print_irqtrace_events(current);
 417                 if (regs)
 418                         show_regs(regs);
 419                 else
 420                         dump_stack();
 421
 422                 if (softlockup_all_cpu_backtrace) {
 423                         /* Avoid generating two back traces for current
 424                          * given that one is already made above
 425                          */
 426                         trigger_allbutself_cpu_backtrace();
 427
 428                         clear_bit(0, &soft_lockup_nmi_warn);
 429                         /* Barrier to sync with other cpus */
 430                         smp_mb__after_atomic();
 431                 }
 432
 433                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 434                 if (softlockup_panic)
 435                         panic("softlockup: hung tasks");
 436                 __this_cpu_write(soft_watchdog_warn, true);
 437         } else
 438                 __this_cpu_write(soft_watchdog_warn, false);
 439
 440         return HRTIMER_RESTART;
 441 }
 442
 443 static void watchdog_enable(unsigned int cpu)
 444 {
 445         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 446         struct completion *done = this_cpu_ptr(&softlockup_completion);
 447
 448         WARN_ON_ONCE(cpu != smp_processor_id());
 449
 450         init_completion(done);
 451         complete(done);
 452
 453         /*
 454          * Start the timer first to prevent the NMI watchdog triggering
 455          * before the timer has a chance to fire.
 456          */
 457         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 458         hrtimer->function = watchdog_timer_fn;
 459         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 460                       HRTIMER_MODE_REL_PINNED_HARD);
 461
 462         /* Initialize timestamp */
 463         __touch_watchdog();
 464         /* Enable the perf event */
 465         if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
 466                 watchdog_nmi_enable(cpu);
 467 }
 468
 469 static void watchdog_disable(unsigned int cpu)
 470 {
 471         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 472
 473         WARN_ON_ONCE(cpu != smp_processor_id());
 474
 475         /*
 476          * Disable the perf event first. That prevents that a large delay
 477          * between disabling the timer and disabling the perf event causes
 478          * the perf NMI to detect a false positive.
 479          */
 480         watchdog_nmi_disable(cpu);
 481         hrtimer_cancel(hrtimer);
 482         wait_for_completion(this_cpu_ptr(&softlockup_completion));
 483 }
 484
 485 static int softlockup_stop_fn(void *data)
 486 {
 487         watchdog_disable(smp_processor_id());
 488         return 0;
 489 }
 490
 491 static void softlockup_stop_all(void)
 492 {
 493         int cpu;
 494
 495         if (!softlockup_initialized)
 496                 return;
 497
 498         for_each_cpu(cpu, &watchdog_allowed_mask)
 499                 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 500
 501         cpumask_clear(&watchdog_allowed_mask);
 502 }
 503
 504 static int softlockup_start_fn(void *data)
 505 {
 506         watchdog_enable(smp_processor_id());
 507         return 0;
 508 }
 509
 510 static void softlockup_start_all(void)
 511 {
 512         int cpu;
 513
 514         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 515         for_each_cpu(cpu, &watchdog_allowed_mask)
 516                 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 517 }
 518
 519 int lockup_detector_online_cpu(unsigned int cpu)
 520 {
 521         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 522                 watchdog_enable(cpu);
 523         return 0;
 524 }
 525
 526 int lockup_detector_offline_cpu(unsigned int cpu)
 527 {
 528         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 529                 watchdog_disable(cpu);
 530         return 0;
 531 }
 532
 533 static void lockup_detector_reconfigure(void)
 534 {
 535         cpus_read_lock();
 536         watchdog_nmi_stop();
 537
 538         softlockup_stop_all();
 539         set_sample_period();
 540         lockup_detector_update_enable();
 541         if (watchdog_enabled && watchdog_thresh)
 542                 softlockup_start_all();
 543
 544         watchdog_nmi_start();
 545         cpus_read_unlock();
 546         /*
 547          * Must be called outside the cpus locked section to prevent
 548          * recursive locking in the perf code.
 549          */
 550         __lockup_detector_cleanup();
 551 }
 552
 553 /*
 554  * Create the watchdog thread infrastructure and configure the detector(s).
 555  *
 556  * The threads are not unparked as watchdog_allowed_mask is empty.  When
 557  * the threads are successfully initialized, take the proper locks and
 558  * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
 559  */
 560 static __init void lockup_detector_setup(void)
 561 {
 562         /*
 563          * If sysctl is off and watchdog got disabled on the command line,
 564          * nothing to do here.
 565          */
 566         lockup_detector_update_enable();
 567
 568         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 569             !(watchdog_enabled && watchdog_thresh))
 570                 return;
 571
 572         mutex_lock(&watchdog_mutex);
 573         lockup_detector_reconfigure();
 574         softlockup_initialized = true;
 575         mutex_unlock(&watchdog_mutex);
 576 }
 577
 578 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 579 static void lockup_detector_reconfigure(void)
 580 {
 581         cpus_read_lock();
 582         watchdog_nmi_stop();
 583         lockup_detector_update_enable();
 584         watchdog_nmi_start();
 585         cpus_read_unlock();
 586 }
 587 static inline void lockup_detector_setup(void)
 588 {
 589         lockup_detector_reconfigure();
 590 }
 591 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 592
 593 static void __lockup_detector_cleanup(void)
 594 {
 595         lockdep_assert_held(&watchdog_mutex);
 596         hardlockup_detector_perf_cleanup();
 597 }
 598
 599 /**
 600  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 601  *
 602  * Caller must not hold the cpu hotplug rwsem.
 603  */
 604 void lockup_detector_cleanup(void)
 605 {
 606         mutex_lock(&watchdog_mutex);
 607         __lockup_detector_cleanup();
 608         mutex_unlock(&watchdog_mutex);
 609 }
 610
 611 /**
 612  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 613  *
 614  * Special interface for parisc. It prevents lockup detector warnings from
 615  * the default pm_poweroff() function which busy loops forever.
 616  */
 617 void lockup_detector_soft_poweroff(void)
 618 {
 619         watchdog_enabled = 0;
 620 }
 621
 622 #ifdef CONFIG_SYSCTL
 623
 624 /* Propagate any changes to the watchdog threads */
 625 static void proc_watchdog_update(void)
 626 {
 627         /* Remove impossible cpus to keep sysctl output clean. */
 628         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 629         lockup_detector_reconfigure();
 630 }
 631
 632 /*
 633  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 634  *
 635  * caller             | table->data points to      | 'which'
 636  * -------------------|----------------------------|--------------------------
 637  * proc_watchdog      | watchdog_user_enabled      | NMI_WATCHDOG_ENABLED |
 638  *                    |                            | SOFT_WATCHDOG_ENABLED
 639  * -------------------|----------------------------|--------------------------
 640  * proc_nmi_watchdog  | nmi_watchdog_user_enabled  | NMI_WATCHDOG_ENABLED
 641  * -------------------|----------------------------|--------------------------
 642  * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
 643  */
 644 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 645                                 void *buffer, size_t *lenp, loff_t *ppos)
 646 {
 647         int err, old, *param = table->data;
 648
 649         mutex_lock(&watchdog_mutex);
 650
 651         if (!write) {
 652                 /*
 653                  * On read synchronize the userspace interface. This is a
 654                  * racy snapshot.
 655                  */
 656                 *param = (watchdog_enabled & which) != 0;
 657                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 658         } else {
 659                 old = READ_ONCE(*param);
 660                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 661                 if (!err && old != READ_ONCE(*param))
 662                         proc_watchdog_update();
 663         }
 664         mutex_unlock(&watchdog_mutex);
 665         return err;
 666 }
 667
 668 /*
 669  * /proc/sys/kernel/watchdog
 670  */
 671 int proc_watchdog(struct ctl_table *table, int write,
 672                   void *buffer, size_t *lenp, loff_t *ppos)
 673 {
 674         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 675                                     table, write, buffer, lenp, ppos);
 676 }
 677
 678 /*
 679  * /proc/sys/kernel/nmi_watchdog
 680  */
 681 int proc_nmi_watchdog(struct ctl_table *table, int write,
 682                       void *buffer, size_t *lenp, loff_t *ppos)
 683 {
 684         if (!nmi_watchdog_available && write)
 685                 return -ENOTSUPP;
 686         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 687                                     table, write, buffer, lenp, ppos);
 688 }
 689
 690 /*
 691  * /proc/sys/kernel/soft_watchdog
 692  */
 693 int proc_soft_watchdog(struct ctl_table *table, int write,
 694                         void *buffer, size_t *lenp, loff_t *ppos)
 695 {
 696         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 697                                     table, write, buffer, lenp, ppos);
 698 }
 699
 700 /*
 701  * /proc/sys/kernel/watchdog_thresh
 702  */
 703 int proc_watchdog_thresh(struct ctl_table *table, int write,
 704                          void *buffer, size_t *lenp, loff_t *ppos)
 705 {
 706         int err, old;
 707
 708         mutex_lock(&watchdog_mutex);
 709
 710         old = READ_ONCE(watchdog_thresh);
 711         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 712
 713         if (!err && write && old != READ_ONCE(watchdog_thresh))
 714                 proc_watchdog_update();
 715
 716         mutex_unlock(&watchdog_mutex);
 717         return err;
 718 }
 719
 720 /*
 721  * The cpumask is the mask of possible cpus that the watchdog can run
 722  * on, not the mask of cpus it is actually running on.  This allows the
 723  * user to specify a mask that will include cpus that have not yet
 724  * been brought online, if desired.
 725  */
 726 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 727                           void *buffer, size_t *lenp, loff_t *ppos)
 728 {
 729         int err;
 730
 731         mutex_lock(&watchdog_mutex);
 732
 733         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 734         if (!err && write)
 735                 proc_watchdog_update();
 736
 737         mutex_unlock(&watchdog_mutex);
 738         return err;
 739 }
 740 #endif /* CONFIG_SYSCTL */
 741
 742 void __init lockup_detector_init(void)
 743 {
 744         if (tick_nohz_full_enabled())
 745                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 746
 747         cpumask_copy(&watchdog_cpumask,
 748                      housekeeping_cpumask(HK_FLAG_TIMER));
 749
 750         if (!watchdog_nmi_probe())
 751                 nmi_watchdog_available = true;
 752         lockup_detector_setup();
 753 }