root/kernel/watchdog.c

DEFINITIONS

1. hardlockup_detector_disable
2. hardlockup_panic_setup
3. hardlockup_all_cpu_backtrace_setup
4. watchdog_nmi_enable
5. watchdog_nmi_disable
6. watchdog_nmi_probe
7. watchdog_nmi_stop
8. watchdog_nmi_start
9. lockup_detector_update_enable
10. softlockup_panic_setup
11. nowatchdog_setup
12. nosoftlockup_setup
13. watchdog_thresh_setup
14. softlockup_all_cpu_backtrace_setup
15. get_softlockup_thresh
16. get_timestamp
17. set_sample_period
18. __touch_watchdog
19. touch_softlockup_watchdog_sched
20. touch_softlockup_watchdog
21. touch_all_softlockup_watchdogs
22. touch_softlockup_watchdog_sync
23. is_softlockup
24. is_hardlockup
25. watchdog_interrupt_count
26. softlockup_fn
27. watchdog_timer_fn
28. watchdog_enable
29. watchdog_disable
30. softlockup_stop_fn
31. softlockup_stop_all
32. softlockup_start_fn
33. softlockup_start_all
34. lockup_detector_online_cpu
35. lockup_detector_offline_cpu
36. lockup_detector_reconfigure
37. lockup_detector_setup
38. lockup_detector_reconfigure
39. lockup_detector_setup
40. __lockup_detector_cleanup
41. lockup_detector_cleanup
42. lockup_detector_soft_poweroff
43. proc_watchdog_update
44. proc_watchdog_common
45. proc_watchdog
46. proc_nmi_watchdog
47. proc_soft_watchdog
48. proc_watchdog_thresh
49. proc_watchdog_cpumask
50. lockup_detector_init

   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 static struct cpumask watchdog_allowed_mask __read_mostly;
  48 
  49 struct cpumask watchdog_cpumask __read_mostly;
  50 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  51 
  52 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  53 /*
  54  * Should we panic when a soft-lockup or hard-lockup occurs:
  55  */
  56 unsigned int __read_mostly hardlockup_panic =
  57                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  58 /*
  59  * We may not want to enable hard lockup detection by default in all cases,
  60  * for example when running the kernel as a guest on a hypervisor. In these
  61  * cases this function can be called to disable hard lockup detection. This
  62  * function should only be executed once by the boot processor before the
  63  * kernel command line parameters are parsed, because otherwise it is not
  64  * possible to override this in hardlockup_panic_setup().
  65  */
  66 void __init hardlockup_detector_disable(void)
  67 {
  68         nmi_watchdog_user_enabled = 0;
  69 }
  70 
  71 static int __init hardlockup_panic_setup(char *str)
  72 {
  73         if (!strncmp(str, "panic", 5))
  74                 hardlockup_panic = 1;
  75         else if (!strncmp(str, "nopanic", 7))
  76                 hardlockup_panic = 0;
  77         else if (!strncmp(str, "0", 1))
  78                 nmi_watchdog_user_enabled = 0;
  79         else if (!strncmp(str, "1", 1))
  80                 nmi_watchdog_user_enabled = 1;
  81         return 1;
  82 }
  83 __setup("nmi_watchdog=", hardlockup_panic_setup);
  84 
  85 # ifdef CONFIG_SMP
  86 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  87 
  88 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
  89 {
  90         sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
  91         return 1;
  92 }
  93 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
  94 # endif /* CONFIG_SMP */
  95 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
  96 
  97 /*
  98  * These functions can be overridden if an architecture implements its
  99  * own hardlockup detector.
 100  *
 101  * watchdog_nmi_enable/disable can be implemented to start and stop when
 102  * softlockup watchdog threads start and stop. The arch must select the
 103  * SOFTLOCKUP_DETECTOR Kconfig.
 104  */
 105 int __weak watchdog_nmi_enable(unsigned int cpu)
 106 {
 107         hardlockup_detector_perf_enable();
 108         return 0;
 109 }
 110 
 111 void __weak watchdog_nmi_disable(unsigned int cpu)
 112 {
 113         hardlockup_detector_perf_disable();
 114 }
 115 
 116 /* Return 0, if a NMI watchdog is available. Error code otherwise */
 117 int __weak __init watchdog_nmi_probe(void)
 118 {
 119         return hardlockup_detector_perf_init();
 120 }
 121 
 122 /**
 123  * watchdog_nmi_stop - Stop the watchdog for reconfiguration
 124  *
 125  * The reconfiguration steps are:
 126  * watchdog_nmi_stop();
 127  * update_variables();
 128  * watchdog_nmi_start();
 129  */
 130 void __weak watchdog_nmi_stop(void) { }
 131 
 132 /**
 133  * watchdog_nmi_start - Start the watchdog after reconfiguration
 134  *
 135  * Counterpart to watchdog_nmi_stop().
 136  *
 137  * The following variables have been updated in update_variables() and
 138  * contain the currently valid configuration:
 139  * - watchdog_enabled
 140  * - watchdog_thresh
 141  * - watchdog_cpumask
 142  */
 143 void __weak watchdog_nmi_start(void) { }
 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 #define SOFTLOCKUP_RESET        ULONG_MAX
 165 
 166 /* Global variables, exported for sysctl */
 167 unsigned int __read_mostly softlockup_panic =
 168                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 169 
 170 static bool softlockup_initialized __read_mostly;
 171 static u64 __read_mostly sample_period;
 172 
 173 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 174 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 175 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 176 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 178 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 179 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 180 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 181 static unsigned long soft_lockup_nmi_warn;
 182 
 183 static int __init softlockup_panic_setup(char *str)
 184 {
 185         softlockup_panic = simple_strtoul(str, NULL, 0);
 186         return 1;
 187 }
 188 __setup("softlockup_panic=", softlockup_panic_setup);
 189 
 190 static int __init nowatchdog_setup(char *str)
 191 {
 192         watchdog_user_enabled = 0;
 193         return 1;
 194 }
 195 __setup("nowatchdog", nowatchdog_setup);
 196 
 197 static int __init nosoftlockup_setup(char *str)
 198 {
 199         soft_watchdog_user_enabled = 0;
 200         return 1;
 201 }
 202 __setup("nosoftlockup", nosoftlockup_setup);
 203 
 204 static int __init watchdog_thresh_setup(char *str)
 205 {
 206         get_option(&str, &watchdog_thresh);
 207         return 1;
 208 }
 209 __setup("watchdog_thresh=", watchdog_thresh_setup);
 210 
 211 #ifdef CONFIG_SMP
 212 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 213 
 214 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 215 {
 216         sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
 217         return 1;
 218 }
 219 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 220 #endif
 221 
 222 static void __lockup_detector_cleanup(void);
 223 
 224 /*
 225  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 226  * lockups can have false positives under extreme conditions. So we generally
 227  * want a higher threshold for soft lockups than for hard lockups. So we couple
 228  * the thresholds with a factor: we make the soft threshold twice the amount of
 229  * time the hard threshold is.
 230  */
 231 static int get_softlockup_thresh(void)
 232 {
 233         return watchdog_thresh * 2;
 234 }
 235 
 236 /*
 237  * Returns seconds, approximately.  We don't need nanosecond
 238  * resolution, and we don't need to waste time with a big divide when
 239  * 2^30ns == 1.074s.
 240  */
 241 static unsigned long get_timestamp(void)
 242 {
 243         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 244 }
 245 
 246 static void set_sample_period(void)
 247 {
 248         /*
 249          * convert watchdog_thresh from seconds to ns
 250          * the divide by 5 is to give hrtimer several chances (two
 251          * or three with the current relation between the soft
 252          * and hard thresholds) to increment before the
 253          * hardlockup detector generates a warning
 254          */
 255         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 256         watchdog_update_hrtimer_threshold(sample_period);
 257 }
 258 
 259 /* Commands for resetting the watchdog */
 260 static void __touch_watchdog(void)
 261 {
 262         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 263 }
 264 
 265 /**
 266  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 267  *
 268  * Call when the scheduler may have stalled for legitimate reasons
 269  * preventing the watchdog task from executing - e.g. the scheduler
 270  * entering idle state.  This should only be used for scheduler events.
 271  * Use touch_softlockup_watchdog() for everything else.
 272  */
 273 notrace void touch_softlockup_watchdog_sched(void)
 274 {
 275         /*
 276          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 277          * gets zeroed here, so use the raw_ operation.
 278          */
 279         raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 280 }
 281 
 282 notrace void touch_softlockup_watchdog(void)
 283 {
 284         touch_softlockup_watchdog_sched();
 285         wq_watchdog_touch(raw_smp_processor_id());
 286 }
 287 EXPORT_SYMBOL(touch_softlockup_watchdog);
 288 
 289 void touch_all_softlockup_watchdogs(void)
 290 {
 291         int cpu;
 292 
 293         /*
 294          * watchdog_mutex cannpt be taken here, as this might be called
 295          * from (soft)interrupt context, so the access to
 296          * watchdog_allowed_cpumask might race with a concurrent update.
 297          *
 298          * The watchdog time stamp can race against a concurrent real
 299          * update as well, the only side effect might be a cycle delay for
 300          * the softlockup check.
 301          */
 302         for_each_cpu(cpu, &watchdog_allowed_mask)
 303                 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
 304         wq_watchdog_touch(-1);
 305 }
 306 
 307 void touch_softlockup_watchdog_sync(void)
 308 {
 309         __this_cpu_write(softlockup_touch_sync, true);
 310         __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 311 }
 312 
 313 static int is_softlockup(unsigned long touch_ts)
 314 {
 315         unsigned long now = get_timestamp();
 316 
 317         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 318                 /* Warn about unreasonable delays. */
 319                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 320                         return now - touch_ts;
 321         }
 322         return 0;
 323 }
 324 
 325 /* watchdog detector functions */
 326 bool is_hardlockup(void)
 327 {
 328         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 329 
 330         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 331                 return true;
 332 
 333         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 334         return false;
 335 }
 336 
 337 static void watchdog_interrupt_count(void)
 338 {
 339         __this_cpu_inc(hrtimer_interrupts);
 340 }
 341 
 342 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 343 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 344 
 345 /*
 346  * The watchdog thread function - touches the timestamp.
 347  *
 348  * It only runs once every sample_period seconds (4 seconds by
 349  * default) to reset the softlockup timestamp. If this gets delayed
 350  * for more than 2*watchdog_thresh seconds then the debug-printout
 351  * triggers in watchdog_timer_fn().
 352  */
 353 static int softlockup_fn(void *data)
 354 {
 355         __this_cpu_write(soft_lockup_hrtimer_cnt,
 356                          __this_cpu_read(hrtimer_interrupts));
 357         __touch_watchdog();
 358         complete(this_cpu_ptr(&softlockup_completion));
 359 
 360         return 0;
 361 }
 362 
 363 /* watchdog kicker functions */
 364 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 365 {
 366         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 367         struct pt_regs *regs = get_irq_regs();
 368         int duration;
 369         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 370 
 371         if (!watchdog_enabled)
 372                 return HRTIMER_NORESTART;
 373 
 374         /* kick the hardlockup detector */
 375         watchdog_interrupt_count();
 376 
 377         /* kick the softlockup detector */
 378         if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 379                 reinit_completion(this_cpu_ptr(&softlockup_completion));
 380                 stop_one_cpu_nowait(smp_processor_id(),
 381                                 softlockup_fn, NULL,
 382                                 this_cpu_ptr(&softlockup_stop_work));
 383         }
 384 
 385         /* .. and repeat */
 386         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 387 
 388         if (touch_ts == SOFTLOCKUP_RESET) {
 389                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 390                         /*
 391                          * If the time stamp was touched atomically
 392                          * make sure the scheduler tick is up to date.
 393                          */
 394                         __this_cpu_write(softlockup_touch_sync, false);
 395                         sched_clock_tick();
 396                 }
 397 
 398                 /* Clear the guest paused flag on watchdog reset */
 399                 kvm_check_and_clear_guest_paused();
 400                 __touch_watchdog();
 401                 return HRTIMER_RESTART;
 402         }
 403 
 404         /* check for a softlockup
 405          * This is done by making sure a high priority task is
 406          * being scheduled.  The task touches the watchdog to
 407          * indicate it is getting cpu time.  If it hasn't then
 408          * this is a good indication some task is hogging the cpu
 409          */
 410         duration = is_softlockup(touch_ts);
 411         if (unlikely(duration)) {
 412                 /*
 413                  * If a virtual machine is stopped by the host it can look to
 414                  * the watchdog like a soft lockup, check to see if the host
 415                  * stopped the vm before we issue the warning
 416                  */
 417                 if (kvm_check_and_clear_guest_paused())
 418                         return HRTIMER_RESTART;
 419 
 420                 /* only warn once */
 421                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 422                         /*
 423                          * When multiple processes are causing softlockups the
 424                          * softlockup detector only warns on the first one
 425                          * because the code relies on a full quiet cycle to
 426                          * re-arm.  The second process prevents the quiet cycle
 427                          * and never gets reported.  Use task pointers to detect
 428                          * this.
 429                          */
 430                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 431                             current) {
 432                                 __this_cpu_write(soft_watchdog_warn, false);
 433                                 __touch_watchdog();
 434                         }
 435                         return HRTIMER_RESTART;
 436                 }
 437 
 438                 if (softlockup_all_cpu_backtrace) {
 439                         /* Prevent multiple soft-lockup reports if one cpu is already
 440                          * engaged in dumping cpu back traces
 441                          */
 442                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 443                                 /* Someone else will report us. Let's give up */
 444                                 __this_cpu_write(soft_watchdog_warn, true);
 445                                 return HRTIMER_RESTART;
 446                         }
 447                 }
 448 
 449                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 450                         smp_processor_id(), duration,
 451                         current->comm, task_pid_nr(current));
 452                 __this_cpu_write(softlockup_task_ptr_saved, current);
 453                 print_modules();
 454                 print_irqtrace_events(current);
 455                 if (regs)
 456                         show_regs(regs);
 457                 else
 458                         dump_stack();
 459 
 460                 if (softlockup_all_cpu_backtrace) {
 461                         /* Avoid generating two back traces for current
 462                          * given that one is already made above
 463                          */
 464                         trigger_allbutself_cpu_backtrace();
 465 
 466                         clear_bit(0, &soft_lockup_nmi_warn);
 467                         /* Barrier to sync with other cpus */
 468                         smp_mb__after_atomic();
 469                 }
 470 
 471                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 472                 if (softlockup_panic)
 473                         panic("softlockup: hung tasks");
 474                 __this_cpu_write(soft_watchdog_warn, true);
 475         } else
 476                 __this_cpu_write(soft_watchdog_warn, false);
 477 
 478         return HRTIMER_RESTART;
 479 }
 480 
 481 static void watchdog_enable(unsigned int cpu)
 482 {
 483         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 484         struct completion *done = this_cpu_ptr(&softlockup_completion);
 485 
 486         WARN_ON_ONCE(cpu != smp_processor_id());
 487 
 488         init_completion(done);
 489         complete(done);
 490 
 491         /*
 492          * Start the timer first to prevent the NMI watchdog triggering
 493          * before the timer has a chance to fire.
 494          */
 495         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 496         hrtimer->function = watchdog_timer_fn;
 497         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 498                       HRTIMER_MODE_REL_PINNED_HARD);
 499 
 500         /* Initialize timestamp */
 501         __touch_watchdog();
 502         /* Enable the perf event */
 503         if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
 504                 watchdog_nmi_enable(cpu);
 505 }
 506 
 507 static void watchdog_disable(unsigned int cpu)
 508 {
 509         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 510 
 511         WARN_ON_ONCE(cpu != smp_processor_id());
 512 
 513         /*
 514          * Disable the perf event first. That prevents that a large delay
 515          * between disabling the timer and disabling the perf event causes
 516          * the perf NMI to detect a false positive.
 517          */
 518         watchdog_nmi_disable(cpu);
 519         hrtimer_cancel(hrtimer);
 520         wait_for_completion(this_cpu_ptr(&softlockup_completion));
 521 }
 522 
 523 static int softlockup_stop_fn(void *data)
 524 {
 525         watchdog_disable(smp_processor_id());
 526         return 0;
 527 }
 528 
 529 static void softlockup_stop_all(void)
 530 {
 531         int cpu;
 532 
 533         if (!softlockup_initialized)
 534                 return;
 535 
 536         for_each_cpu(cpu, &watchdog_allowed_mask)
 537                 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 538 
 539         cpumask_clear(&watchdog_allowed_mask);
 540 }
 541 
 542 static int softlockup_start_fn(void *data)
 543 {
 544         watchdog_enable(smp_processor_id());
 545         return 0;
 546 }
 547 
 548 static void softlockup_start_all(void)
 549 {
 550         int cpu;
 551 
 552         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 553         for_each_cpu(cpu, &watchdog_allowed_mask)
 554                 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 555 }
 556 
 557 int lockup_detector_online_cpu(unsigned int cpu)
 558 {
 559         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 560                 watchdog_enable(cpu);
 561         return 0;
 562 }
 563 
 564 int lockup_detector_offline_cpu(unsigned int cpu)
 565 {
 566         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 567                 watchdog_disable(cpu);
 568         return 0;
 569 }
 570 
 571 static void lockup_detector_reconfigure(void)
 572 {
 573         cpus_read_lock();
 574         watchdog_nmi_stop();
 575 
 576         softlockup_stop_all();
 577         set_sample_period();
 578         lockup_detector_update_enable();
 579         if (watchdog_enabled && watchdog_thresh)
 580                 softlockup_start_all();
 581 
 582         watchdog_nmi_start();
 583         cpus_read_unlock();
 584         /*
 585          * Must be called outside the cpus locked section to prevent
 586          * recursive locking in the perf code.
 587          */
 588         __lockup_detector_cleanup();
 589 }
 590 
 591 /*
 592  * Create the watchdog thread infrastructure and configure the detector(s).
 593  *
 594  * The threads are not unparked as watchdog_allowed_mask is empty.  When
 595  * the threads are successfully initialized, take the proper locks and
 596  * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
 597  */
 598 static __init void lockup_detector_setup(void)
 599 {
 600         /*
 601          * If sysctl is off and watchdog got disabled on the command line,
 602          * nothing to do here.
 603          */
 604         lockup_detector_update_enable();
 605 
 606         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 607             !(watchdog_enabled && watchdog_thresh))
 608                 return;
 609 
 610         mutex_lock(&watchdog_mutex);
 611         lockup_detector_reconfigure();
 612         softlockup_initialized = true;
 613         mutex_unlock(&watchdog_mutex);
 614 }
 615 
 616 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 617 static void lockup_detector_reconfigure(void)
 618 {
 619         cpus_read_lock();
 620         watchdog_nmi_stop();
 621         lockup_detector_update_enable();
 622         watchdog_nmi_start();
 623         cpus_read_unlock();
 624 }
 625 static inline void lockup_detector_setup(void)
 626 {
 627         lockup_detector_reconfigure();
 628 }
 629 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 630 
 631 static void __lockup_detector_cleanup(void)
 632 {
 633         lockdep_assert_held(&watchdog_mutex);
 634         hardlockup_detector_perf_cleanup();
 635 }
 636 
 637 /**
 638  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 639  *
 640  * Caller must not hold the cpu hotplug rwsem.
 641  */
 642 void lockup_detector_cleanup(void)
 643 {
 644         mutex_lock(&watchdog_mutex);
 645         __lockup_detector_cleanup();
 646         mutex_unlock(&watchdog_mutex);
 647 }
 648 
 649 /**
 650  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 651  *
 652  * Special interface for parisc. It prevents lockup detector warnings from
 653  * the default pm_poweroff() function which busy loops forever.
 654  */
 655 void lockup_detector_soft_poweroff(void)
 656 {
 657         watchdog_enabled = 0;
 658 }
 659 
 660 #ifdef CONFIG_SYSCTL
 661 
 662 /* Propagate any changes to the watchdog threads */
 663 static void proc_watchdog_update(void)
 664 {
 665         /* Remove impossible cpus to keep sysctl output clean. */
 666         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 667         lockup_detector_reconfigure();
 668 }
 669 
 670 /*
 671  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 672  *
 673  * caller             | table->data points to      | 'which'
 674  * -------------------|----------------------------|--------------------------
 675  * proc_watchdog      | watchdog_user_enabled      | NMI_WATCHDOG_ENABLED |
 676  *                    |                            | SOFT_WATCHDOG_ENABLED
 677  * -------------------|----------------------------|--------------------------
 678  * proc_nmi_watchdog  | nmi_watchdog_user_enabled  | NMI_WATCHDOG_ENABLED
 679  * -------------------|----------------------------|--------------------------
 680  * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
 681  */
 682 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 683                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 684 {
 685         int err, old, *param = table->data;
 686 
 687         mutex_lock(&watchdog_mutex);
 688 
 689         if (!write) {
 690                 /*
 691                  * On read synchronize the userspace interface. This is a
 692                  * racy snapshot.
 693                  */
 694                 *param = (watchdog_enabled & which) != 0;
 695                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 696         } else {
 697                 old = READ_ONCE(*param);
 698                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 699                 if (!err && old != READ_ONCE(*param))
 700                         proc_watchdog_update();
 701         }
 702         mutex_unlock(&watchdog_mutex);
 703         return err;
 704 }
 705 
 706 /*
 707  * /proc/sys/kernel/watchdog
 708  */
 709 int proc_watchdog(struct ctl_table *table, int write,
 710                   void __user *buffer, size_t *lenp, loff_t *ppos)
 711 {
 712         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 713                                     table, write, buffer, lenp, ppos);
 714 }
 715 
 716 /*
 717  * /proc/sys/kernel/nmi_watchdog
 718  */
 719 int proc_nmi_watchdog(struct ctl_table *table, int write,
 720                       void __user *buffer, size_t *lenp, loff_t *ppos)
 721 {
 722         if (!nmi_watchdog_available && write)
 723                 return -ENOTSUPP;
 724         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 725                                     table, write, buffer, lenp, ppos);
 726 }
 727 
 728 /*
 729  * /proc/sys/kernel/soft_watchdog
 730  */
 731 int proc_soft_watchdog(struct ctl_table *table, int write,
 732                         void __user *buffer, size_t *lenp, loff_t *ppos)
 733 {
 734         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 735                                     table, write, buffer, lenp, ppos);
 736 }
 737 
 738 /*
 739  * /proc/sys/kernel/watchdog_thresh
 740  */
 741 int proc_watchdog_thresh(struct ctl_table *table, int write,
 742                          void __user *buffer, size_t *lenp, loff_t *ppos)
 743 {
 744         int err, old;
 745 
 746         mutex_lock(&watchdog_mutex);
 747 
 748         old = READ_ONCE(watchdog_thresh);
 749         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 750 
 751         if (!err && write && old != READ_ONCE(watchdog_thresh))
 752                 proc_watchdog_update();
 753 
 754         mutex_unlock(&watchdog_mutex);
 755         return err;
 756 }
 757 
 758 /*
 759  * The cpumask is the mask of possible cpus that the watchdog can run
 760  * on, not the mask of cpus it is actually running on.  This allows the
 761  * user to specify a mask that will include cpus that have not yet
 762  * been brought online, if desired.
 763  */
 764 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 765                           void __user *buffer, size_t *lenp, loff_t *ppos)
 766 {
 767         int err;
 768 
 769         mutex_lock(&watchdog_mutex);
 770 
 771         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 772         if (!err && write)
 773                 proc_watchdog_update();
 774 
 775         mutex_unlock(&watchdog_mutex);
 776         return err;
 777 }
 778 #endif /* CONFIG_SYSCTL */
 779 
 780 void __init lockup_detector_init(void)
 781 {
 782         if (tick_nohz_full_enabled())
 783                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 784 
 785         cpumask_copy(&watchdog_cpumask,
 786                      housekeeping_cpumask(HK_FLAG_TIMER));
 787 
 788         if (!watchdog_nmi_probe())
 789                 nmi_watchdog_available = true;
 790         lockup_detector_setup();
 791 }