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                         IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
  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 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 /*
 158  * Delay the soflockup report when running a known slow code.
 159  * It does _not_ affect the timestamp of the last successdul reschedule.
 160  */
 161 #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
 162
 163 #ifdef CONFIG_SMP
 164 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 165 #endif
 166
 167 static struct cpumask watchdog_allowed_mask __read_mostly;
 168
 169 /* Global variables, exported for sysctl */
 170 unsigned int __read_mostly softlockup_panic =
 171                         IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
 172
 173 static bool softlockup_initialized __read_mostly;
 174 static u64 __read_mostly sample_period;
 175
 176 /* Timestamp taken after the last successful reschedule. */
 177 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 178 /* Timestamp of the last softlockup report. */
 179 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
 180 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 181 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 182 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 183 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 184 static unsigned long soft_lockup_nmi_warn;
 185
 186 static int __init softlockup_panic_setup(char *str)
 187 {
 188         softlockup_panic = simple_strtoul(str, NULL, 0);
 189         return 1;
 190 }
 191 __setup("softlockup_panic=", softlockup_panic_setup);
 192
 193 static int __init nowatchdog_setup(char *str)
 194 {
 195         watchdog_user_enabled = 0;
 196         return 1;
 197 }
 198 __setup("nowatchdog", nowatchdog_setup);
 199
 200 static int __init nosoftlockup_setup(char *str)
 201 {
 202         soft_watchdog_user_enabled = 0;
 203         return 1;
 204 }
 205 __setup("nosoftlockup", nosoftlockup_setup);
 206
 207 static int __init watchdog_thresh_setup(char *str)
 208 {
 209         get_option(&str, &watchdog_thresh);
 210         return 1;
 211 }
 212 __setup("watchdog_thresh=", watchdog_thresh_setup);
 213
 214 static void __lockup_detector_cleanup(void);
 215
 216 /*
 217  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 218  * lockups can have false positives under extreme conditions. So we generally
 219  * want a higher threshold for soft lockups than for hard lockups. So we couple
 220  * the thresholds with a factor: we make the soft threshold twice the amount of
 221  * time the hard threshold is.
 222  */
 223 static int get_softlockup_thresh(void)
 224 {
 225         return watchdog_thresh * 2;
 226 }
 227
 228 /*
 229  * Returns seconds, approximately.  We don't need nanosecond
 230  * resolution, and we don't need to waste time with a big divide when
 231  * 2^30ns == 1.074s.
 232  */
 233 static unsigned long get_timestamp(void)
 234 {
 235         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 236 }
 237
 238 static void set_sample_period(void)
 239 {
 240         /*
 241          * convert watchdog_thresh from seconds to ns
 242          * the divide by 5 is to give hrtimer several chances (two
 243          * or three with the current relation between the soft
 244          * and hard thresholds) to increment before the
 245          * hardlockup detector generates a warning
 246          */
 247         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 248         watchdog_update_hrtimer_threshold(sample_period);
 249 }
 250
 251 static void update_report_ts(void)
 252 {
 253         __this_cpu_write(watchdog_report_ts, get_timestamp());
 254 }
 255
 256 /* Commands for resetting the watchdog */
 257 static void update_touch_ts(void)
 258 {
 259         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 260         update_report_ts();
 261 }
 262
 263 /**
 264  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 265  *
 266  * Call when the scheduler may have stalled for legitimate reasons
 267  * preventing the watchdog task from executing - e.g. the scheduler
 268  * entering idle state.  This should only be used for scheduler events.
 269  * Use touch_softlockup_watchdog() for everything else.
 270  */
 271 notrace void touch_softlockup_watchdog_sched(void)
 272 {
 273         /*
 274          * Preemption can be enabled.  It doesn't matter which CPU's watchdog
 275          * report period gets restarted here, so use the raw_ operation.
 276          */
 277         raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 278 }
 279
 280 notrace void touch_softlockup_watchdog(void)
 281 {
 282         touch_softlockup_watchdog_sched();
 283         wq_watchdog_touch(raw_smp_processor_id());
 284 }
 285 EXPORT_SYMBOL(touch_softlockup_watchdog);
 286
 287 void touch_all_softlockup_watchdogs(void)
 288 {
 289         int cpu;
 290
 291         /*
 292          * watchdog_mutex cannpt be taken here, as this might be called
 293          * from (soft)interrupt context, so the access to
 294          * watchdog_allowed_cpumask might race with a concurrent update.
 295          *
 296          * The watchdog time stamp can race against a concurrent real
 297          * update as well, the only side effect might be a cycle delay for
 298          * the softlockup check.
 299          */
 300         for_each_cpu(cpu, &watchdog_allowed_mask) {
 301                 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
 302                 wq_watchdog_touch(cpu);
 303         }
 304 }
 305
 306 void touch_softlockup_watchdog_sync(void)
 307 {
 308         __this_cpu_write(softlockup_touch_sync, true);
 309         __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 310 }
 311
 312 static int is_softlockup(unsigned long touch_ts,
 313                          unsigned long period_ts,
 314                          unsigned long now)
 315 {
 316         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 317                 /* Warn about unreasonable delays. */
 318                 if (time_after(now, period_ts + get_softlockup_thresh()))
 319                         return now - touch_ts;
 320         }
 321         return 0;
 322 }
 323
 324 /* watchdog detector functions */
 325 bool is_hardlockup(void)
 326 {
 327         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 328
 329         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 330                 return true;
 331
 332         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 333         return false;
 334 }
 335
 336 static void watchdog_interrupt_count(void)
 337 {
 338         __this_cpu_inc(hrtimer_interrupts);
 339 }
 340
 341 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 342 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 343
 344 /*
 345  * The watchdog feed function - touches the timestamp.
 346  *
 347  * It only runs once every sample_period seconds (4 seconds by
 348  * default) to reset the softlockup timestamp. If this gets delayed
 349  * for more than 2*watchdog_thresh seconds then the debug-printout
 350  * triggers in watchdog_timer_fn().
 351  */
 352 static int softlockup_fn(void *data)
 353 {
 354         update_touch_ts();
 355         complete(this_cpu_ptr(&softlockup_completion));
 356
 357         return 0;
 358 }
 359
 360 /* watchdog kicker functions */
 361 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 362 {
 363         unsigned long touch_ts, period_ts, now;
 364         struct pt_regs *regs = get_irq_regs();
 365         int duration;
 366         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 367
 368         if (!watchdog_enabled)
 369                 return HRTIMER_NORESTART;
 370
 371         /* kick the hardlockup detector */
 372         watchdog_interrupt_count();
 373
 374         /* kick the softlockup detector */
 375         if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 376                 reinit_completion(this_cpu_ptr(&softlockup_completion));
 377                 stop_one_cpu_nowait(smp_processor_id(),
 378                                 softlockup_fn, NULL,
 379                                 this_cpu_ptr(&softlockup_stop_work));
 380         }
 381
 382         /* .. and repeat */
 383         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 384
 385         /*
 386          * Read the current timestamp first. It might become invalid anytime
 387          * when a virtual machine is stopped by the host or when the watchog
 388          * is touched from NMI.
 389          */
 390         now = get_timestamp();
 391         /*
 392          * If a virtual machine is stopped by the host it can look to
 393          * the watchdog like a soft lockup. This function touches the watchdog.
 394          */
 395         kvm_check_and_clear_guest_paused();
 396         /*
 397          * The stored timestamp is comparable with @now only when not touched.
 398          * It might get touched anytime from NMI. Make sure that is_softlockup()
 399          * uses the same (valid) value.
 400          */
 401         period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
 402
 403         /* Reset the interval when touched by known problematic code. */
 404         if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
 405                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 406                         /*
 407                          * If the time stamp was touched atomically
 408                          * make sure the scheduler tick is up to date.
 409                          */
 410                         __this_cpu_write(softlockup_touch_sync, false);
 411                         sched_clock_tick();
 412                 }
 413
 414                 update_report_ts();
 415                 return HRTIMER_RESTART;
 416         }
 417
 418         /* Check for a softlockup. */
 419         touch_ts = __this_cpu_read(watchdog_touch_ts);
 420         duration = is_softlockup(touch_ts, period_ts, now);
 421         if (unlikely(duration)) {
 422                 /*
 423                  * Prevent multiple soft-lockup reports if one cpu is already
 424                  * engaged in dumping all cpu back traces.
 425                  */
 426                 if (softlockup_all_cpu_backtrace) {
 427                         if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
 428                                 return HRTIMER_RESTART;
 429                 }
 430
 431                 /* Start period for the next softlockup warning. */
 432                 update_report_ts();
 433
 434                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 435                         smp_processor_id(), duration,
 436                         current->comm, task_pid_nr(current));
 437                 print_modules();
 438                 print_irqtrace_events(current);
 439                 if (regs)
 440                         show_regs(regs);
 441                 else
 442                         dump_stack();
 443
 444                 if (softlockup_all_cpu_backtrace) {
 445                         trigger_allbutself_cpu_backtrace();
 446                         clear_bit_unlock(0, &soft_lockup_nmi_warn);
 447                 }
 448
 449                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 450                 if (softlockup_panic)
 451                         panic("softlockup: hung tasks");
 452         }
 453
 454         return HRTIMER_RESTART;
 455 }
 456
 457 static void watchdog_enable(unsigned int cpu)
 458 {
 459         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 460         struct completion *done = this_cpu_ptr(&softlockup_completion);
 461
 462         WARN_ON_ONCE(cpu != smp_processor_id());
 463
 464         init_completion(done);
 465         complete(done);
 466
 467         /*
 468          * Start the timer first to prevent the NMI watchdog triggering
 469          * before the timer has a chance to fire.
 470          */
 471         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 472         hrtimer->function = watchdog_timer_fn;
 473         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 474                       HRTIMER_MODE_REL_PINNED_HARD);
 475
 476         /* Initialize timestamp */
 477         update_touch_ts();
 478         /* Enable the perf event */
 479         if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
 480                 watchdog_nmi_enable(cpu);
 481 }
 482
 483 static void watchdog_disable(unsigned int cpu)
 484 {
 485         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 486
 487         WARN_ON_ONCE(cpu != smp_processor_id());
 488
 489         /*
 490          * Disable the perf event first. That prevents that a large delay
 491          * between disabling the timer and disabling the perf event causes
 492          * the perf NMI to detect a false positive.
 493          */
 494         watchdog_nmi_disable(cpu);
 495         hrtimer_cancel(hrtimer);
 496         wait_for_completion(this_cpu_ptr(&softlockup_completion));
 497 }
 498
 499 static int softlockup_stop_fn(void *data)
 500 {
 501         watchdog_disable(smp_processor_id());
 502         return 0;
 503 }
 504
 505 static void softlockup_stop_all(void)
 506 {
 507         int cpu;
 508
 509         if (!softlockup_initialized)
 510                 return;
 511
 512         for_each_cpu(cpu, &watchdog_allowed_mask)
 513                 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 514
 515         cpumask_clear(&watchdog_allowed_mask);
 516 }
 517
 518 static int softlockup_start_fn(void *data)
 519 {
 520         watchdog_enable(smp_processor_id());
 521         return 0;
 522 }
 523
 524 static void softlockup_start_all(void)
 525 {
 526         int cpu;
 527
 528         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 529         for_each_cpu(cpu, &watchdog_allowed_mask)
 530                 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 531 }
 532
 533 int lockup_detector_online_cpu(unsigned int cpu)
 534 {
 535         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 536                 watchdog_enable(cpu);
 537         return 0;
 538 }
 539
 540 int lockup_detector_offline_cpu(unsigned int cpu)
 541 {
 542         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 543                 watchdog_disable(cpu);
 544         return 0;
 545 }
 546
 547 static void __lockup_detector_reconfigure(void)
 548 {
 549         cpus_read_lock();
 550         watchdog_nmi_stop();
 551
 552         softlockup_stop_all();
 553         set_sample_period();
 554         lockup_detector_update_enable();
 555         if (watchdog_enabled && watchdog_thresh)
 556                 softlockup_start_all();
 557
 558         watchdog_nmi_start();
 559         cpus_read_unlock();
 560         /*
 561          * Must be called outside the cpus locked section to prevent
 562          * recursive locking in the perf code.
 563          */
 564         __lockup_detector_cleanup();
 565 }
 566
 567 void lockup_detector_reconfigure(void)
 568 {
 569         mutex_lock(&watchdog_mutex);
 570         __lockup_detector_reconfigure();
 571         mutex_unlock(&watchdog_mutex);
 572 }
 573
 574 /*
 575  * Create the watchdog infrastructure and configure the detector(s).
 576  */
 577 static __init void lockup_detector_setup(void)
 578 {
 579         /*
 580          * If sysctl is off and watchdog got disabled on the command line,
 581          * nothing to do here.
 582          */
 583         lockup_detector_update_enable();
 584
 585         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 586             !(watchdog_enabled && watchdog_thresh))
 587                 return;
 588
 589         mutex_lock(&watchdog_mutex);
 590         __lockup_detector_reconfigure();
 591         softlockup_initialized = true;
 592         mutex_unlock(&watchdog_mutex);
 593 }
 594
 595 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 596 static void __lockup_detector_reconfigure(void)
 597 {
 598         cpus_read_lock();
 599         watchdog_nmi_stop();
 600         lockup_detector_update_enable();
 601         watchdog_nmi_start();
 602         cpus_read_unlock();
 603 }
 604 void lockup_detector_reconfigure(void)
 605 {
 606         __lockup_detector_reconfigure();
 607 }
 608 static inline void lockup_detector_setup(void)
 609 {
 610         __lockup_detector_reconfigure();
 611 }
 612 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 613
 614 static void __lockup_detector_cleanup(void)
 615 {
 616         lockdep_assert_held(&watchdog_mutex);
 617         hardlockup_detector_perf_cleanup();
 618 }
 619
 620 /**
 621  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 622  *
 623  * Caller must not hold the cpu hotplug rwsem.
 624  */
 625 void lockup_detector_cleanup(void)
 626 {
 627         mutex_lock(&watchdog_mutex);
 628         __lockup_detector_cleanup();
 629         mutex_unlock(&watchdog_mutex);
 630 }
 631
 632 /**
 633  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 634  *
 635  * Special interface for parisc. It prevents lockup detector warnings from
 636  * the default pm_poweroff() function which busy loops forever.
 637  */
 638 void lockup_detector_soft_poweroff(void)
 639 {
 640         watchdog_enabled = 0;
 641 }
 642
 643 #ifdef CONFIG_SYSCTL
 644
 645 /* Propagate any changes to the watchdog infrastructure */
 646 static void proc_watchdog_update(void)
 647 {
 648         /* Remove impossible cpus to keep sysctl output clean. */
 649         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 650         __lockup_detector_reconfigure();
 651 }
 652
 653 /*
 654  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 655  *
 656  * caller             | table->data points to      | 'which'
 657  * -------------------|----------------------------|--------------------------
 658  * proc_watchdog      | watchdog_user_enabled      | NMI_WATCHDOG_ENABLED |
 659  *                    |                            | SOFT_WATCHDOG_ENABLED
 660  * -------------------|----------------------------|--------------------------
 661  * proc_nmi_watchdog  | nmi_watchdog_user_enabled  | NMI_WATCHDOG_ENABLED
 662  * -------------------|----------------------------|--------------------------
 663  * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
 664  */
 665 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 666                                 void *buffer, size_t *lenp, loff_t *ppos)
 667 {
 668         int err, old, *param = table->data;
 669
 670         mutex_lock(&watchdog_mutex);
 671
 672         if (!write) {
 673                 /*
 674                  * On read synchronize the userspace interface. This is a
 675                  * racy snapshot.
 676                  */
 677                 *param = (watchdog_enabled & which) != 0;
 678                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 679         } else {
 680                 old = READ_ONCE(*param);
 681                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 682                 if (!err && old != READ_ONCE(*param))
 683                         proc_watchdog_update();
 684         }
 685         mutex_unlock(&watchdog_mutex);
 686         return err;
 687 }
 688
 689 /*
 690  * /proc/sys/kernel/watchdog
 691  */
 692 int proc_watchdog(struct ctl_table *table, int write,
 693                   void *buffer, size_t *lenp, loff_t *ppos)
 694 {
 695         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 696                                     table, write, buffer, lenp, ppos);
 697 }
 698
 699 /*
 700  * /proc/sys/kernel/nmi_watchdog
 701  */
 702 int proc_nmi_watchdog(struct ctl_table *table, int write,
 703                       void *buffer, size_t *lenp, loff_t *ppos)
 704 {
 705         if (!nmi_watchdog_available && write)
 706                 return -ENOTSUPP;
 707         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 708                                     table, write, buffer, lenp, ppos);
 709 }
 710
 711 /*
 712  * /proc/sys/kernel/soft_watchdog
 713  */
 714 int proc_soft_watchdog(struct ctl_table *table, int write,
 715                         void *buffer, size_t *lenp, loff_t *ppos)
 716 {
 717         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 718                                     table, write, buffer, lenp, ppos);
 719 }
 720
 721 /*
 722  * /proc/sys/kernel/watchdog_thresh
 723  */
 724 int proc_watchdog_thresh(struct ctl_table *table, int write,
 725                          void *buffer, size_t *lenp, loff_t *ppos)
 726 {
 727         int err, old;
 728
 729         mutex_lock(&watchdog_mutex);
 730
 731         old = READ_ONCE(watchdog_thresh);
 732         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 733
 734         if (!err && write && old != READ_ONCE(watchdog_thresh))
 735                 proc_watchdog_update();
 736
 737         mutex_unlock(&watchdog_mutex);
 738         return err;
 739 }
 740
 741 /*
 742  * The cpumask is the mask of possible cpus that the watchdog can run
 743  * on, not the mask of cpus it is actually running on.  This allows the
 744  * user to specify a mask that will include cpus that have not yet
 745  * been brought online, if desired.
 746  */
 747 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 748                           void *buffer, size_t *lenp, loff_t *ppos)
 749 {
 750         int err;
 751
 752         mutex_lock(&watchdog_mutex);
 753
 754         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 755         if (!err && write)
 756                 proc_watchdog_update();
 757
 758         mutex_unlock(&watchdog_mutex);
 759         return err;
 760 }
 761
 762 static const int sixty = 60;
 763
 764 static struct ctl_table watchdog_sysctls[] = {
 765         {
 766                 .procname       = "watchdog",
 767                 .data           = &watchdog_user_enabled,
 768                 .maxlen         = sizeof(int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_watchdog,
 771                 .extra1         = SYSCTL_ZERO,
 772                 .extra2         = SYSCTL_ONE,
 773         },
 774         {
 775                 .procname       = "watchdog_thresh",
 776                 .data           = &watchdog_thresh,
 777                 .maxlen         = sizeof(int),
 778                 .mode           = 0644,
 779                 .proc_handler   = proc_watchdog_thresh,
 780                 .extra1         = SYSCTL_ZERO,
 781                 .extra2         = (void *)&sixty,
 782         },
 783         {
 784                 .procname       = "nmi_watchdog",
 785                 .data           = &nmi_watchdog_user_enabled,
 786                 .maxlen         = sizeof(int),
 787                 .mode           = NMI_WATCHDOG_SYSCTL_PERM,
 788                 .proc_handler   = proc_nmi_watchdog,
 789                 .extra1         = SYSCTL_ZERO,
 790                 .extra2         = SYSCTL_ONE,
 791         },
 792         {
 793                 .procname       = "watchdog_cpumask",
 794                 .data           = &watchdog_cpumask_bits,
 795                 .maxlen         = NR_CPUS,
 796                 .mode           = 0644,
 797                 .proc_handler   = proc_watchdog_cpumask,
 798         },
 799 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 800         {
 801                 .procname       = "soft_watchdog",
 802                 .data           = &soft_watchdog_user_enabled,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0644,
 805                 .proc_handler   = proc_soft_watchdog,
 806                 .extra1         = SYSCTL_ZERO,
 807                 .extra2         = SYSCTL_ONE,
 808         },
 809         {
 810                 .procname       = "softlockup_panic",
 811                 .data           = &softlockup_panic,
 812                 .maxlen         = sizeof(int),
 813                 .mode           = 0644,
 814                 .proc_handler   = proc_dointvec_minmax,
 815                 .extra1         = SYSCTL_ZERO,
 816                 .extra2         = SYSCTL_ONE,
 817         },
 818 #ifdef CONFIG_SMP
 819         {
 820                 .procname       = "softlockup_all_cpu_backtrace",
 821                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 822                 .maxlen         = sizeof(int),
 823                 .mode           = 0644,
 824                 .proc_handler   = proc_dointvec_minmax,
 825                 .extra1         = SYSCTL_ZERO,
 826                 .extra2         = SYSCTL_ONE,
 827         },
 828 #endif /* CONFIG_SMP */
 829 #endif
 830 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 831         {
 832                 .procname       = "hardlockup_panic",
 833                 .data           = &hardlockup_panic,
 834                 .maxlen         = sizeof(int),
 835                 .mode           = 0644,
 836                 .proc_handler   = proc_dointvec_minmax,
 837                 .extra1         = SYSCTL_ZERO,
 838                 .extra2         = SYSCTL_ONE,
 839         },
 840 #ifdef CONFIG_SMP
 841         {
 842                 .procname       = "hardlockup_all_cpu_backtrace",
 843                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 844                 .maxlen         = sizeof(int),
 845                 .mode           = 0644,
 846                 .proc_handler   = proc_dointvec_minmax,
 847                 .extra1         = SYSCTL_ZERO,
 848                 .extra2         = SYSCTL_ONE,
 849         },
 850 #endif /* CONFIG_SMP */
 851 #endif
 852         {}
 853 };
 854
 855 static void __init watchdog_sysctl_init(void)
 856 {
 857         register_sysctl_init("kernel", watchdog_sysctls);
 858 }
 859 #else
 860 #define watchdog_sysctl_init() do { } while (0)
 861 #endif /* CONFIG_SYSCTL */
 862
 863 void __init lockup_detector_init(void)
 864 {
 865         if (tick_nohz_full_enabled())
 866                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 867
 868         cpumask_copy(&watchdog_cpumask,
 869                      housekeeping_cpumask(HK_TYPE_TIMER));
 870
 871         if (!watchdog_nmi_probe())
 872                 nmi_watchdog_available = true;
 873         lockup_detector_setup();
 874         watchdog_sysctl_init();
 875 }