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_HARDLOCKUP_DETECTOR_SPARC64)
  33 # define WATCHDOG_HARDLOCKUP_DEFAULT    1
  34 #else
  35 # define WATCHDOG_HARDLOCKUP_DEFAULT    0
  36 #endif
  37
  38 unsigned long __read_mostly watchdog_enabled;
  39 int __read_mostly watchdog_user_enabled = 1;
  40 static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT;
  41 static int __read_mostly watchdog_softlockup_user_enabled = 1;
  42 int __read_mostly watchdog_thresh = 10;
  43 static int __read_mostly watchdog_hardlockup_available;
  44
  45 struct cpumask watchdog_cpumask __read_mostly;
  46 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  47
  48 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  49
  50 # ifdef CONFIG_SMP
  51 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  52 # endif /* CONFIG_SMP */
  53
  54 /*
  55  * Should we panic when a soft-lockup or hard-lockup occurs:
  56  */
  57 unsigned int __read_mostly hardlockup_panic =
  58                         IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
  59 /*
  60  * We may not want to enable hard lockup detection by default in all cases,
  61  * for example when running the kernel as a guest on a hypervisor. In these
  62  * cases this function can be called to disable hard lockup detection. This
  63  * function should only be executed once by the boot processor before the
  64  * kernel command line parameters are parsed, because otherwise it is not
  65  * possible to override this in hardlockup_panic_setup().
  66  */
  67 void __init hardlockup_detector_disable(void)
  68 {
  69         watchdog_hardlockup_user_enabled = 0;
  70 }
  71
  72 static int __init hardlockup_panic_setup(char *str)
  73 {
  74         if (!strncmp(str, "panic", 5))
  75                 hardlockup_panic = 1;
  76         else if (!strncmp(str, "nopanic", 7))
  77                 hardlockup_panic = 0;
  78         else if (!strncmp(str, "0", 1))
  79                 watchdog_hardlockup_user_enabled = 0;
  80         else if (!strncmp(str, "1", 1))
  81                 watchdog_hardlockup_user_enabled = 1;
  82         return 1;
  83 }
  84 __setup("nmi_watchdog=", hardlockup_panic_setup);
  85
  86 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
  87
  88 #if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER)
  89
  90 static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts);
  91 static DEFINE_PER_CPU(int, hrtimer_interrupts_saved);
  92 static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned);
  93 static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched);
  94 static unsigned long watchdog_hardlockup_all_cpu_dumped;
  95
  96 notrace void arch_touch_nmi_watchdog(void)
  97 {
  98         /*
  99          * Using __raw here because some code paths have
 100          * preemption enabled.  If preemption is enabled
 101          * then interrupts should be enabled too, in which
 102          * case we shouldn't have to worry about the watchdog
 103          * going off.
 104          */
 105         raw_cpu_write(watchdog_hardlockup_touched, true);
 106 }
 107 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 108
 109 void watchdog_hardlockup_touch_cpu(unsigned int cpu)
 110 {
 111         per_cpu(watchdog_hardlockup_touched, cpu) = true;
 112 }
 113
 114 static bool is_hardlockup(unsigned int cpu)
 115 {
 116         int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu));
 117
 118         if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
 119                 return true;
 120
 121         /*
 122          * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE
 123          * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is
 124          * written/read by a single CPU.
 125          */
 126         per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
 127
 128         return false;
 129 }
 130
 131 static void watchdog_hardlockup_kick(void)
 132 {
 133         int new_interrupts;
 134
 135         new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts));
 136         watchdog_buddy_check_hardlockup(new_interrupts);
 137 }
 138
 139 void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 140 {
 141         if (per_cpu(watchdog_hardlockup_touched, cpu)) {
 142                 per_cpu(watchdog_hardlockup_touched, cpu) = false;
 143                 return;
 144         }
 145
 146         /*
 147          * Check for a hardlockup by making sure the CPU's timer
 148          * interrupt is incrementing. The timer interrupt should have
 149          * fired multiple times before we overflow'd. If it hasn't
 150          * then this is a good indication the cpu is stuck
 151          */
 152         if (is_hardlockup(cpu)) {
 153                 unsigned int this_cpu = smp_processor_id();
 154                 struct cpumask backtrace_mask;
 155
 156                 cpumask_copy(&backtrace_mask, cpu_online_mask);
 157
 158                 /* Only print hardlockups once. */
 159                 if (per_cpu(watchdog_hardlockup_warned, cpu))
 160                         return;
 161
 162                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
 163                 print_modules();
 164                 print_irqtrace_events(current);
 165                 if (cpu == this_cpu) {
 166                         if (regs)
 167                                 show_regs(regs);
 168                         else
 169                                 dump_stack();
 170                         cpumask_clear_cpu(cpu, &backtrace_mask);
 171                 } else {
 172                         if (trigger_single_cpu_backtrace(cpu))
 173                                 cpumask_clear_cpu(cpu, &backtrace_mask);
 174                 }
 175
 176                 /*
 177                  * Perform multi-CPU dump only once to avoid multiple
 178                  * hardlockups generating interleaving traces
 179                  */
 180                 if (sysctl_hardlockup_all_cpu_backtrace &&
 181                     !test_and_set_bit(0, &watchdog_hardlockup_all_cpu_dumped))
 182                         trigger_cpumask_backtrace(&backtrace_mask);
 183
 184                 if (hardlockup_panic)
 185                         nmi_panic(regs, "Hard LOCKUP");
 186
 187                 per_cpu(watchdog_hardlockup_warned, cpu) = true;
 188         } else {
 189                 per_cpu(watchdog_hardlockup_warned, cpu) = false;
 190         }
 191 }
 192
 193 #else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
 194
 195 static inline void watchdog_hardlockup_kick(void) { }
 196
 197 #endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
 198
 199 /*
 200  * These functions can be overridden based on the configured hardlockdup detector.
 201  *
 202  * watchdog_hardlockup_enable/disable can be implemented to start and stop when
 203  * softlockup watchdog start and stop. The detector must select the
 204  * SOFTLOCKUP_DETECTOR Kconfig.
 205  */
 206 void __weak watchdog_hardlockup_enable(unsigned int cpu) { }
 207
 208 void __weak watchdog_hardlockup_disable(unsigned int cpu) { }
 209
 210 /*
 211  * Watchdog-detector specific API.
 212  *
 213  * Return 0 when hardlockup watchdog is available, negative value otherwise.
 214  * Note that the negative value means that a delayed probe might
 215  * succeed later.
 216  */
 217 int __weak __init watchdog_hardlockup_probe(void)
 218 {
 219         return -ENODEV;
 220 }
 221
 222 /**
 223  * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration
 224  *
 225  * The reconfiguration steps are:
 226  * watchdog_hardlockup_stop();
 227  * update_variables();
 228  * watchdog_hardlockup_start();
 229  */
 230 void __weak watchdog_hardlockup_stop(void) { }
 231
 232 /**
 233  * watchdog_hardlockup_start - Start the watchdog after reconfiguration
 234  *
 235  * Counterpart to watchdog_hardlockup_stop().
 236  *
 237  * The following variables have been updated in update_variables() and
 238  * contain the currently valid configuration:
 239  * - watchdog_enabled
 240  * - watchdog_thresh
 241  * - watchdog_cpumask
 242  */
 243 void __weak watchdog_hardlockup_start(void) { }
 244
 245 /**
 246  * lockup_detector_update_enable - Update the sysctl enable bit
 247  *
 248  * Caller needs to make sure that the hard watchdogs are off, so this
 249  * can't race with watchdog_hardlockup_disable().
 250  */
 251 static void lockup_detector_update_enable(void)
 252 {
 253         watchdog_enabled = 0;
 254         if (!watchdog_user_enabled)
 255                 return;
 256         if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled)
 257                 watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED;
 258         if (watchdog_softlockup_user_enabled)
 259                 watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED;
 260 }
 261
 262 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 263
 264 /*
 265  * Delay the soflockup report when running a known slow code.
 266  * It does _not_ affect the timestamp of the last successdul reschedule.
 267  */
 268 #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
 269
 270 #ifdef CONFIG_SMP
 271 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 272 #endif
 273
 274 static struct cpumask watchdog_allowed_mask __read_mostly;
 275
 276 /* Global variables, exported for sysctl */
 277 unsigned int __read_mostly softlockup_panic =
 278                         IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
 279
 280 static bool softlockup_initialized __read_mostly;
 281 static u64 __read_mostly sample_period;
 282
 283 /* Timestamp taken after the last successful reschedule. */
 284 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 285 /* Timestamp of the last softlockup report. */
 286 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
 287 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 288 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 289 static unsigned long soft_lockup_nmi_warn;
 290
 291 static int __init nowatchdog_setup(char *str)
 292 {
 293         watchdog_user_enabled = 0;
 294         return 1;
 295 }
 296 __setup("nowatchdog", nowatchdog_setup);
 297
 298 static int __init nosoftlockup_setup(char *str)
 299 {
 300         watchdog_softlockup_user_enabled = 0;
 301         return 1;
 302 }
 303 __setup("nosoftlockup", nosoftlockup_setup);
 304
 305 static int __init watchdog_thresh_setup(char *str)
 306 {
 307         get_option(&str, &watchdog_thresh);
 308         return 1;
 309 }
 310 __setup("watchdog_thresh=", watchdog_thresh_setup);
 311
 312 static void __lockup_detector_cleanup(void);
 313
 314 /*
 315  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 316  * lockups can have false positives under extreme conditions. So we generally
 317  * want a higher threshold for soft lockups than for hard lockups. So we couple
 318  * the thresholds with a factor: we make the soft threshold twice the amount of
 319  * time the hard threshold is.
 320  */
 321 static int get_softlockup_thresh(void)
 322 {
 323         return watchdog_thresh * 2;
 324 }
 325
 326 /*
 327  * Returns seconds, approximately.  We don't need nanosecond
 328  * resolution, and we don't need to waste time with a big divide when
 329  * 2^30ns == 1.074s.
 330  */
 331 static unsigned long get_timestamp(void)
 332 {
 333         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 334 }
 335
 336 static void set_sample_period(void)
 337 {
 338         /*
 339          * convert watchdog_thresh from seconds to ns
 340          * the divide by 5 is to give hrtimer several chances (two
 341          * or three with the current relation between the soft
 342          * and hard thresholds) to increment before the
 343          * hardlockup detector generates a warning
 344          */
 345         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 346         watchdog_update_hrtimer_threshold(sample_period);
 347 }
 348
 349 static void update_report_ts(void)
 350 {
 351         __this_cpu_write(watchdog_report_ts, get_timestamp());
 352 }
 353
 354 /* Commands for resetting the watchdog */
 355 static void update_touch_ts(void)
 356 {
 357         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 358         update_report_ts();
 359 }
 360
 361 /**
 362  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 363  *
 364  * Call when the scheduler may have stalled for legitimate reasons
 365  * preventing the watchdog task from executing - e.g. the scheduler
 366  * entering idle state.  This should only be used for scheduler events.
 367  * Use touch_softlockup_watchdog() for everything else.
 368  */
 369 notrace void touch_softlockup_watchdog_sched(void)
 370 {
 371         /*
 372          * Preemption can be enabled.  It doesn't matter which CPU's watchdog
 373          * report period gets restarted here, so use the raw_ operation.
 374          */
 375         raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 376 }
 377
 378 notrace void touch_softlockup_watchdog(void)
 379 {
 380         touch_softlockup_watchdog_sched();
 381         wq_watchdog_touch(raw_smp_processor_id());
 382 }
 383 EXPORT_SYMBOL(touch_softlockup_watchdog);
 384
 385 void touch_all_softlockup_watchdogs(void)
 386 {
 387         int cpu;
 388
 389         /*
 390          * watchdog_mutex cannpt be taken here, as this might be called
 391          * from (soft)interrupt context, so the access to
 392          * watchdog_allowed_cpumask might race with a concurrent update.
 393          *
 394          * The watchdog time stamp can race against a concurrent real
 395          * update as well, the only side effect might be a cycle delay for
 396          * the softlockup check.
 397          */
 398         for_each_cpu(cpu, &watchdog_allowed_mask) {
 399                 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
 400                 wq_watchdog_touch(cpu);
 401         }
 402 }
 403
 404 void touch_softlockup_watchdog_sync(void)
 405 {
 406         __this_cpu_write(softlockup_touch_sync, true);
 407         __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 408 }
 409
 410 static int is_softlockup(unsigned long touch_ts,
 411                          unsigned long period_ts,
 412                          unsigned long now)
 413 {
 414         if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
 415                 /* Warn about unreasonable delays. */
 416                 if (time_after(now, period_ts + get_softlockup_thresh()))
 417                         return now - touch_ts;
 418         }
 419         return 0;
 420 }
 421
 422 /* watchdog detector functions */
 423 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 424 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 425
 426 /*
 427  * The watchdog feed function - touches the timestamp.
 428  *
 429  * It only runs once every sample_period seconds (4 seconds by
 430  * default) to reset the softlockup timestamp. If this gets delayed
 431  * for more than 2*watchdog_thresh seconds then the debug-printout
 432  * triggers in watchdog_timer_fn().
 433  */
 434 static int softlockup_fn(void *data)
 435 {
 436         update_touch_ts();
 437         complete(this_cpu_ptr(&softlockup_completion));
 438
 439         return 0;
 440 }
 441
 442 /* watchdog kicker functions */
 443 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 444 {
 445         unsigned long touch_ts, period_ts, now;
 446         struct pt_regs *regs = get_irq_regs();
 447         int duration;
 448         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 449
 450         if (!watchdog_enabled)
 451                 return HRTIMER_NORESTART;
 452
 453         watchdog_hardlockup_kick();
 454
 455         /* kick the softlockup detector */
 456         if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 457                 reinit_completion(this_cpu_ptr(&softlockup_completion));
 458                 stop_one_cpu_nowait(smp_processor_id(),
 459                                 softlockup_fn, NULL,
 460                                 this_cpu_ptr(&softlockup_stop_work));
 461         }
 462
 463         /* .. and repeat */
 464         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 465
 466         /*
 467          * Read the current timestamp first. It might become invalid anytime
 468          * when a virtual machine is stopped by the host or when the watchog
 469          * is touched from NMI.
 470          */
 471         now = get_timestamp();
 472         /*
 473          * If a virtual machine is stopped by the host it can look to
 474          * the watchdog like a soft lockup. This function touches the watchdog.
 475          */
 476         kvm_check_and_clear_guest_paused();
 477         /*
 478          * The stored timestamp is comparable with @now only when not touched.
 479          * It might get touched anytime from NMI. Make sure that is_softlockup()
 480          * uses the same (valid) value.
 481          */
 482         period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
 483
 484         /* Reset the interval when touched by known problematic code. */
 485         if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
 486                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 487                         /*
 488                          * If the time stamp was touched atomically
 489                          * make sure the scheduler tick is up to date.
 490                          */
 491                         __this_cpu_write(softlockup_touch_sync, false);
 492                         sched_clock_tick();
 493                 }
 494
 495                 update_report_ts();
 496                 return HRTIMER_RESTART;
 497         }
 498
 499         /* Check for a softlockup. */
 500         touch_ts = __this_cpu_read(watchdog_touch_ts);
 501         duration = is_softlockup(touch_ts, period_ts, now);
 502         if (unlikely(duration)) {
 503                 /*
 504                  * Prevent multiple soft-lockup reports if one cpu is already
 505                  * engaged in dumping all cpu back traces.
 506                  */
 507                 if (softlockup_all_cpu_backtrace) {
 508                         if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
 509                                 return HRTIMER_RESTART;
 510                 }
 511
 512                 /* Start period for the next softlockup warning. */
 513                 update_report_ts();
 514
 515                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 516                         smp_processor_id(), duration,
 517                         current->comm, task_pid_nr(current));
 518                 print_modules();
 519                 print_irqtrace_events(current);
 520                 if (regs)
 521                         show_regs(regs);
 522                 else
 523                         dump_stack();
 524
 525                 if (softlockup_all_cpu_backtrace) {
 526                         trigger_allbutself_cpu_backtrace();
 527                         clear_bit_unlock(0, &soft_lockup_nmi_warn);
 528                 }
 529
 530                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 531                 if (softlockup_panic)
 532                         panic("softlockup: hung tasks");
 533         }
 534
 535         return HRTIMER_RESTART;
 536 }
 537
 538 static void watchdog_enable(unsigned int cpu)
 539 {
 540         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 541         struct completion *done = this_cpu_ptr(&softlockup_completion);
 542
 543         WARN_ON_ONCE(cpu != smp_processor_id());
 544
 545         init_completion(done);
 546         complete(done);
 547
 548         /*
 549          * Start the timer first to prevent the hardlockup watchdog triggering
 550          * before the timer has a chance to fire.
 551          */
 552         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 553         hrtimer->function = watchdog_timer_fn;
 554         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 555                       HRTIMER_MODE_REL_PINNED_HARD);
 556
 557         /* Initialize timestamp */
 558         update_touch_ts();
 559         /* Enable the hardlockup detector */
 560         if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)
 561                 watchdog_hardlockup_enable(cpu);
 562 }
 563
 564 static void watchdog_disable(unsigned int cpu)
 565 {
 566         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 567
 568         WARN_ON_ONCE(cpu != smp_processor_id());
 569
 570         /*
 571          * Disable the hardlockup detector first. That prevents that a large
 572          * delay between disabling the timer and disabling the hardlockup
 573          * detector causes a false positive.
 574          */
 575         watchdog_hardlockup_disable(cpu);
 576         hrtimer_cancel(hrtimer);
 577         wait_for_completion(this_cpu_ptr(&softlockup_completion));
 578 }
 579
 580 static int softlockup_stop_fn(void *data)
 581 {
 582         watchdog_disable(smp_processor_id());
 583         return 0;
 584 }
 585
 586 static void softlockup_stop_all(void)
 587 {
 588         int cpu;
 589
 590         if (!softlockup_initialized)
 591                 return;
 592
 593         for_each_cpu(cpu, &watchdog_allowed_mask)
 594                 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 595
 596         cpumask_clear(&watchdog_allowed_mask);
 597 }
 598
 599 static int softlockup_start_fn(void *data)
 600 {
 601         watchdog_enable(smp_processor_id());
 602         return 0;
 603 }
 604
 605 static void softlockup_start_all(void)
 606 {
 607         int cpu;
 608
 609         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 610         for_each_cpu(cpu, &watchdog_allowed_mask)
 611                 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 612 }
 613
 614 int lockup_detector_online_cpu(unsigned int cpu)
 615 {
 616         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 617                 watchdog_enable(cpu);
 618         return 0;
 619 }
 620
 621 int lockup_detector_offline_cpu(unsigned int cpu)
 622 {
 623         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 624                 watchdog_disable(cpu);
 625         return 0;
 626 }
 627
 628 static void __lockup_detector_reconfigure(void)
 629 {
 630         cpus_read_lock();
 631         watchdog_hardlockup_stop();
 632
 633         softlockup_stop_all();
 634         set_sample_period();
 635         lockup_detector_update_enable();
 636         if (watchdog_enabled && watchdog_thresh)
 637                 softlockup_start_all();
 638
 639         watchdog_hardlockup_start();
 640         cpus_read_unlock();
 641         /*
 642          * Must be called outside the cpus locked section to prevent
 643          * recursive locking in the perf code.
 644          */
 645         __lockup_detector_cleanup();
 646 }
 647
 648 void lockup_detector_reconfigure(void)
 649 {
 650         mutex_lock(&watchdog_mutex);
 651         __lockup_detector_reconfigure();
 652         mutex_unlock(&watchdog_mutex);
 653 }
 654
 655 /*
 656  * Create the watchdog infrastructure and configure the detector(s).
 657  */
 658 static __init void lockup_detector_setup(void)
 659 {
 660         /*
 661          * If sysctl is off and watchdog got disabled on the command line,
 662          * nothing to do here.
 663          */
 664         lockup_detector_update_enable();
 665
 666         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 667             !(watchdog_enabled && watchdog_thresh))
 668                 return;
 669
 670         mutex_lock(&watchdog_mutex);
 671         __lockup_detector_reconfigure();
 672         softlockup_initialized = true;
 673         mutex_unlock(&watchdog_mutex);
 674 }
 675
 676 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 677 static void __lockup_detector_reconfigure(void)
 678 {
 679         cpus_read_lock();
 680         watchdog_hardlockup_stop();
 681         lockup_detector_update_enable();
 682         watchdog_hardlockup_start();
 683         cpus_read_unlock();
 684 }
 685 void lockup_detector_reconfigure(void)
 686 {
 687         __lockup_detector_reconfigure();
 688 }
 689 static inline void lockup_detector_setup(void)
 690 {
 691         __lockup_detector_reconfigure();
 692 }
 693 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 694
 695 static void __lockup_detector_cleanup(void)
 696 {
 697         lockdep_assert_held(&watchdog_mutex);
 698         hardlockup_detector_perf_cleanup();
 699 }
 700
 701 /**
 702  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 703  *
 704  * Caller must not hold the cpu hotplug rwsem.
 705  */
 706 void lockup_detector_cleanup(void)
 707 {
 708         mutex_lock(&watchdog_mutex);
 709         __lockup_detector_cleanup();
 710         mutex_unlock(&watchdog_mutex);
 711 }
 712
 713 /**
 714  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 715  *
 716  * Special interface for parisc. It prevents lockup detector warnings from
 717  * the default pm_poweroff() function which busy loops forever.
 718  */
 719 void lockup_detector_soft_poweroff(void)
 720 {
 721         watchdog_enabled = 0;
 722 }
 723
 724 #ifdef CONFIG_SYSCTL
 725
 726 /* Propagate any changes to the watchdog infrastructure */
 727 static void proc_watchdog_update(void)
 728 {
 729         /* Remove impossible cpus to keep sysctl output clean. */
 730         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 731         __lockup_detector_reconfigure();
 732 }
 733
 734 /*
 735  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 736  *
 737  * caller             | table->data points to            | 'which'
 738  * -------------------|----------------------------------|-------------------------------
 739  * proc_watchdog      | watchdog_user_enabled            | WATCHDOG_HARDLOCKUP_ENABLED |
 740  *                    |                                  | WATCHDOG_SOFTOCKUP_ENABLED
 741  * -------------------|----------------------------------|-------------------------------
 742  * proc_nmi_watchdog  | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED
 743  * -------------------|----------------------------------|-------------------------------
 744  * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED
 745  */
 746 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 747                                 void *buffer, size_t *lenp, loff_t *ppos)
 748 {
 749         int err, old, *param = table->data;
 750
 751         mutex_lock(&watchdog_mutex);
 752
 753         if (!write) {
 754                 /*
 755                  * On read synchronize the userspace interface. This is a
 756                  * racy snapshot.
 757                  */
 758                 *param = (watchdog_enabled & which) != 0;
 759                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 760         } else {
 761                 old = READ_ONCE(*param);
 762                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 763                 if (!err && old != READ_ONCE(*param))
 764                         proc_watchdog_update();
 765         }
 766         mutex_unlock(&watchdog_mutex);
 767         return err;
 768 }
 769
 770 /*
 771  * /proc/sys/kernel/watchdog
 772  */
 773 int proc_watchdog(struct ctl_table *table, int write,
 774                   void *buffer, size_t *lenp, loff_t *ppos)
 775 {
 776         return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED |
 777                                     WATCHDOG_SOFTOCKUP_ENABLED,
 778                                     table, write, buffer, lenp, ppos);
 779 }
 780
 781 /*
 782  * /proc/sys/kernel/nmi_watchdog
 783  */
 784 int proc_nmi_watchdog(struct ctl_table *table, int write,
 785                       void *buffer, size_t *lenp, loff_t *ppos)
 786 {
 787         if (!watchdog_hardlockup_available && write)
 788                 return -ENOTSUPP;
 789         return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED,
 790                                     table, write, buffer, lenp, ppos);
 791 }
 792
 793 /*
 794  * /proc/sys/kernel/soft_watchdog
 795  */
 796 int proc_soft_watchdog(struct ctl_table *table, int write,
 797                         void *buffer, size_t *lenp, loff_t *ppos)
 798 {
 799         return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED,
 800                                     table, write, buffer, lenp, ppos);
 801 }
 802
 803 /*
 804  * /proc/sys/kernel/watchdog_thresh
 805  */
 806 int proc_watchdog_thresh(struct ctl_table *table, int write,
 807                          void *buffer, size_t *lenp, loff_t *ppos)
 808 {
 809         int err, old;
 810
 811         mutex_lock(&watchdog_mutex);
 812
 813         old = READ_ONCE(watchdog_thresh);
 814         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 815
 816         if (!err && write && old != READ_ONCE(watchdog_thresh))
 817                 proc_watchdog_update();
 818
 819         mutex_unlock(&watchdog_mutex);
 820         return err;
 821 }
 822
 823 /*
 824  * The cpumask is the mask of possible cpus that the watchdog can run
 825  * on, not the mask of cpus it is actually running on.  This allows the
 826  * user to specify a mask that will include cpus that have not yet
 827  * been brought online, if desired.
 828  */
 829 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 830                           void *buffer, size_t *lenp, loff_t *ppos)
 831 {
 832         int err;
 833
 834         mutex_lock(&watchdog_mutex);
 835
 836         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 837         if (!err && write)
 838                 proc_watchdog_update();
 839
 840         mutex_unlock(&watchdog_mutex);
 841         return err;
 842 }
 843
 844 static const int sixty = 60;
 845
 846 static struct ctl_table watchdog_sysctls[] = {
 847         {
 848                 .procname       = "watchdog",
 849                 .data           = &watchdog_user_enabled,
 850                 .maxlen         = sizeof(int),
 851                 .mode           = 0644,
 852                 .proc_handler   = proc_watchdog,
 853                 .extra1         = SYSCTL_ZERO,
 854                 .extra2         = SYSCTL_ONE,
 855         },
 856         {
 857                 .procname       = "watchdog_thresh",
 858                 .data           = &watchdog_thresh,
 859                 .maxlen         = sizeof(int),
 860                 .mode           = 0644,
 861                 .proc_handler   = proc_watchdog_thresh,
 862                 .extra1         = SYSCTL_ZERO,
 863                 .extra2         = (void *)&sixty,
 864         },
 865         {
 866                 .procname       = "watchdog_cpumask",
 867                 .data           = &watchdog_cpumask_bits,
 868                 .maxlen         = NR_CPUS,
 869                 .mode           = 0644,
 870                 .proc_handler   = proc_watchdog_cpumask,
 871         },
 872 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 873         {
 874                 .procname       = "soft_watchdog",
 875                 .data           = &watchdog_softlockup_user_enabled,
 876                 .maxlen         = sizeof(int),
 877                 .mode           = 0644,
 878                 .proc_handler   = proc_soft_watchdog,
 879                 .extra1         = SYSCTL_ZERO,
 880                 .extra2         = SYSCTL_ONE,
 881         },
 882         {
 883                 .procname       = "softlockup_panic",
 884                 .data           = &softlockup_panic,
 885                 .maxlen         = sizeof(int),
 886                 .mode           = 0644,
 887                 .proc_handler   = proc_dointvec_minmax,
 888                 .extra1         = SYSCTL_ZERO,
 889                 .extra2         = SYSCTL_ONE,
 890         },
 891 #ifdef CONFIG_SMP
 892         {
 893                 .procname       = "softlockup_all_cpu_backtrace",
 894                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = proc_dointvec_minmax,
 898                 .extra1         = SYSCTL_ZERO,
 899                 .extra2         = SYSCTL_ONE,
 900         },
 901 #endif /* CONFIG_SMP */
 902 #endif
 903 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 904         {
 905                 .procname       = "hardlockup_panic",
 906                 .data           = &hardlockup_panic,
 907                 .maxlen         = sizeof(int),
 908                 .mode           = 0644,
 909                 .proc_handler   = proc_dointvec_minmax,
 910                 .extra1         = SYSCTL_ZERO,
 911                 .extra2         = SYSCTL_ONE,
 912         },
 913 #ifdef CONFIG_SMP
 914         {
 915                 .procname       = "hardlockup_all_cpu_backtrace",
 916                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = proc_dointvec_minmax,
 920                 .extra1         = SYSCTL_ZERO,
 921                 .extra2         = SYSCTL_ONE,
 922         },
 923 #endif /* CONFIG_SMP */
 924 #endif
 925         {}
 926 };
 927
 928 static struct ctl_table watchdog_hardlockup_sysctl[] = {
 929         {
 930                 .procname       = "nmi_watchdog",
 931                 .data           = &watchdog_hardlockup_user_enabled,
 932                 .maxlen         = sizeof(int),
 933                 .mode           = 0444,
 934                 .proc_handler   = proc_nmi_watchdog,
 935                 .extra1         = SYSCTL_ZERO,
 936                 .extra2         = SYSCTL_ONE,
 937         },
 938         {}
 939 };
 940
 941 static void __init watchdog_sysctl_init(void)
 942 {
 943         register_sysctl_init("kernel", watchdog_sysctls);
 944
 945         if (watchdog_hardlockup_available)
 946                 watchdog_hardlockup_sysctl[0].mode = 0644;
 947         register_sysctl_init("kernel", watchdog_hardlockup_sysctl);
 948 }
 949
 950 #else
 951 #define watchdog_sysctl_init() do { } while (0)
 952 #endif /* CONFIG_SYSCTL */
 953
 954 static void __init lockup_detector_delay_init(struct work_struct *work);
 955 static bool allow_lockup_detector_init_retry __initdata;
 956
 957 static struct work_struct detector_work __initdata =
 958                 __WORK_INITIALIZER(detector_work, lockup_detector_delay_init);
 959
 960 static void __init lockup_detector_delay_init(struct work_struct *work)
 961 {
 962         int ret;
 963
 964         ret = watchdog_hardlockup_probe();
 965         if (ret) {
 966                 pr_info("Delayed init of the lockup detector failed: %d\n", ret);
 967                 pr_info("Hard watchdog permanently disabled\n");
 968                 return;
 969         }
 970
 971         allow_lockup_detector_init_retry = false;
 972
 973         watchdog_hardlockup_available = true;
 974         lockup_detector_setup();
 975 }
 976
 977 /*
 978  * lockup_detector_retry_init - retry init lockup detector if possible.
 979  *
 980  * Retry hardlockup detector init. It is useful when it requires some
 981  * functionality that has to be initialized later on a particular
 982  * platform.
 983  */
 984 void __init lockup_detector_retry_init(void)
 985 {
 986         /* Must be called before late init calls */
 987         if (!allow_lockup_detector_init_retry)
 988                 return;
 989
 990         schedule_work(&detector_work);
 991 }
 992
 993 /*
 994  * Ensure that optional delayed hardlockup init is proceed before
 995  * the init code and memory is freed.
 996  */
 997 static int __init lockup_detector_check(void)
 998 {
 999         /* Prevent any later retry. */
1000         allow_lockup_detector_init_retry = false;
1001
1002         /* Make sure no work is pending. */
1003         flush_work(&detector_work);
1004
1005         watchdog_sysctl_init();
1006
1007         return 0;
1008
1009 }
1010 late_initcall_sync(lockup_detector_check);
1011
1012 void __init lockup_detector_init(void)
1013 {
1014         if (tick_nohz_full_enabled())
1015                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1016
1017         cpumask_copy(&watchdog_cpumask,
1018                      housekeeping_cpumask(HK_TYPE_TIMER));
1019
1020         if (!watchdog_hardlockup_probe())
1021                 watchdog_hardlockup_available = true;
1022         else
1023                 allow_lockup_detector_init_retry = true;
1024
1025         lockup_detector_setup();
1026 }