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