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 int __read_mostly nmi_watchdog_available;
  46
  47 struct cpumask watchdog_allowed_mask __read_mostly;
  48
  49 struct cpumask watchdog_cpumask __read_mostly;
  50 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  51
  52 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  53 /*
  54  * Should we panic when a soft-lockup or hard-lockup occurs:
  55  */
  56 unsigned int __read_mostly hardlockup_panic =
  57                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  58 /*
  59  * We may not want to enable hard lockup detection by default in all cases,
  60  * for example when running the kernel as a guest on a hypervisor. In these
  61  * cases this function can be called to disable hard lockup detection. This
  62  * function should only be executed once by the boot processor before the
  63  * kernel command line parameters are parsed, because otherwise it is not
  64  * possible to override this in hardlockup_panic_setup().
  65  */
  66 void __init hardlockup_detector_disable(void)
  67 {
  68         nmi_watchdog_user_enabled = 0;
  69 }
  70
  71 static int __init hardlockup_panic_setup(char *str)
  72 {
  73         if (!strncmp(str, "panic", 5))
  74                 hardlockup_panic = 1;
  75         else if (!strncmp(str, "nopanic", 7))
  76                 hardlockup_panic = 0;
  77         else if (!strncmp(str, "0", 1))
  78                 nmi_watchdog_user_enabled = 0;
  79         else if (!strncmp(str, "1", 1))
  80                 nmi_watchdog_user_enabled = 1;
  81         return 1;
  82 }
  83 __setup("nmi_watchdog=", hardlockup_panic_setup);
  84
  85 # ifdef CONFIG_SMP
  86 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  87
  88 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
  89 {
  90         sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
  91         return 1;
  92 }
  93 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
  94 # endif /* CONFIG_SMP */
  95 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
  96
  97 /*
  98  * These functions can be overridden if an architecture implements its
  99  * own hardlockup detector.
 100  *
 101  * watchdog_nmi_enable/disable can be implemented to start and stop when
 102  * softlockup watchdog threads start and stop. The arch must select the
 103  * SOFTLOCKUP_DETECTOR Kconfig.
 104  */
 105 int __weak watchdog_nmi_enable(unsigned int cpu)
 106 {
 107         hardlockup_detector_perf_enable();
 108         return 0;
 109 }
 110
 111 void __weak watchdog_nmi_disable(unsigned int cpu)
 112 {
 113         hardlockup_detector_perf_disable();
 114 }
 115
 116 /* Return 0, if a NMI watchdog is available. Error code otherwise */
 117 int __weak __init watchdog_nmi_probe(void)
 118 {
 119         return hardlockup_detector_perf_init();
 120 }
 121
 122 /**
 123  * watchdog_nmi_stop - Stop the watchdog for reconfiguration
 124  *
 125  * The reconfiguration steps are:
 126  * watchdog_nmi_stop();
 127  * update_variables();
 128  * watchdog_nmi_start();
 129  */
 130 void __weak watchdog_nmi_stop(void) { }
 131
 132 /**
 133  * watchdog_nmi_start - Start the watchdog after reconfiguration
 134  *
 135  * Counterpart to watchdog_nmi_stop().
 136  *
 137  * The following variables have been updated in update_variables() and
 138  * contain the currently valid configuration:
 139  * - watchdog_enabled
 140  * - watchdog_thresh
 141  * - watchdog_cpumask
 142  */
 143 void __weak watchdog_nmi_start(void) { }
 144
 145 /**
 146  * lockup_detector_update_enable - Update the sysctl enable bit
 147  *
 148  * Caller needs to make sure that the NMI/perf watchdogs are off, so this
 149  * can't race with watchdog_nmi_disable().
 150  */
 151 static void lockup_detector_update_enable(void)
 152 {
 153         watchdog_enabled = 0;
 154         if (!watchdog_user_enabled)
 155                 return;
 156         if (nmi_watchdog_available && nmi_watchdog_user_enabled)
 157                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 158         if (soft_watchdog_user_enabled)
 159                 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
 160 }
 161
 162 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 163
 164 #define SOFTLOCKUP_RESET        ULONG_MAX
 165
 166 /* Global variables, exported for sysctl */
 167 unsigned int __read_mostly softlockup_panic =
 168                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 169
 170 static bool softlockup_initialized __read_mostly;
 171 static u64 __read_mostly sample_period;
 172
 173 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 174 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 175 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 176 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 178 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 179 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 180 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 181 static unsigned long soft_lockup_nmi_warn;
 182
 183 static int __init softlockup_panic_setup(char *str)
 184 {
 185         softlockup_panic = simple_strtoul(str, NULL, 0);
 186         return 1;
 187 }
 188 __setup("softlockup_panic=", softlockup_panic_setup);
 189
 190 static int __init nowatchdog_setup(char *str)
 191 {
 192         watchdog_user_enabled = 0;
 193         return 1;
 194 }
 195 __setup("nowatchdog", nowatchdog_setup);
 196
 197 static int __init nosoftlockup_setup(char *str)
 198 {
 199         soft_watchdog_user_enabled = 0;
 200         return 1;
 201 }
 202 __setup("nosoftlockup", nosoftlockup_setup);
 203
 204 #ifdef CONFIG_SMP
 205 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 206
 207 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 208 {
 209         sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
 210         return 1;
 211 }
 212 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 213 #endif
 214
 215 static void __lockup_detector_cleanup(void);
 216
 217 /*
 218  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 219  * lockups can have false positives under extreme conditions. So we generally
 220  * want a higher threshold for soft lockups than for hard lockups. So we couple
 221  * the thresholds with a factor: we make the soft threshold twice the amount of
 222  * time the hard threshold is.
 223  */
 224 static int get_softlockup_thresh(void)
 225 {
 226         return watchdog_thresh * 2;
 227 }
 228
 229 /*
 230  * Returns seconds, approximately.  We don't need nanosecond
 231  * resolution, and we don't need to waste time with a big divide when
 232  * 2^30ns == 1.074s.
 233  */
 234 static unsigned long get_timestamp(void)
 235 {
 236         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 237 }
 238
 239 static void set_sample_period(void)
 240 {
 241         /*
 242          * convert watchdog_thresh from seconds to ns
 243          * the divide by 5 is to give hrtimer several chances (two
 244          * or three with the current relation between the soft
 245          * and hard thresholds) to increment before the
 246          * hardlockup detector generates a warning
 247          */
 248         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 249         watchdog_update_hrtimer_threshold(sample_period);
 250 }
 251
 252 /* Commands for resetting the watchdog */
 253 static void __touch_watchdog(void)
 254 {
 255         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 256 }
 257
 258 /**
 259  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 260  *
 261  * Call when the scheduler may have stalled for legitimate reasons
 262  * preventing the watchdog task from executing - e.g. the scheduler
 263  * entering idle state.  This should only be used for scheduler events.
 264  * Use touch_softlockup_watchdog() for everything else.
 265  */
 266 notrace void touch_softlockup_watchdog_sched(void)
 267 {
 268         /*
 269          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 270          * gets zeroed here, so use the raw_ operation.
 271          */
 272         raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 273 }
 274
 275 notrace void touch_softlockup_watchdog(void)
 276 {
 277         touch_softlockup_watchdog_sched();
 278         wq_watchdog_touch(raw_smp_processor_id());
 279 }
 280 EXPORT_SYMBOL(touch_softlockup_watchdog);
 281
 282 void touch_all_softlockup_watchdogs(void)
 283 {
 284         int cpu;
 285
 286         /*
 287          * watchdog_mutex cannpt be taken here, as this might be called
 288          * from (soft)interrupt context, so the access to
 289          * watchdog_allowed_cpumask might race with a concurrent update.
 290          *
 291          * The watchdog time stamp can race against a concurrent real
 292          * update as well, the only side effect might be a cycle delay for
 293          * the softlockup check.
 294          */
 295         for_each_cpu(cpu, &watchdog_allowed_mask)
 296                 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
 297         wq_watchdog_touch(-1);
 298 }
 299
 300 void touch_softlockup_watchdog_sync(void)
 301 {
 302         __this_cpu_write(softlockup_touch_sync, true);
 303         __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
 304 }
 305
 306 static int is_softlockup(unsigned long touch_ts)
 307 {
 308         unsigned long now = get_timestamp();
 309
 310         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 311                 /* Warn about unreasonable delays. */
 312                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 313                         return now - touch_ts;
 314         }
 315         return 0;
 316 }
 317
 318 /* watchdog detector functions */
 319 bool is_hardlockup(void)
 320 {
 321         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 322
 323         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 324                 return true;
 325
 326         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 327         return false;
 328 }
 329
 330 static void watchdog_interrupt_count(void)
 331 {
 332         __this_cpu_inc(hrtimer_interrupts);
 333 }
 334
 335 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 336 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 337
 338 /*
 339  * The watchdog thread function - touches the timestamp.
 340  *
 341  * It only runs once every sample_period seconds (4 seconds by
 342  * default) to reset the softlockup timestamp. If this gets delayed
 343  * for more than 2*watchdog_thresh seconds then the debug-printout
 344  * triggers in watchdog_timer_fn().
 345  */
 346 static int softlockup_fn(void *data)
 347 {
 348         __this_cpu_write(soft_lockup_hrtimer_cnt,
 349                          __this_cpu_read(hrtimer_interrupts));
 350         __touch_watchdog();
 351         complete(this_cpu_ptr(&softlockup_completion));
 352
 353         return 0;
 354 }
 355
 356 /* watchdog kicker functions */
 357 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 358 {
 359         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 360         struct pt_regs *regs = get_irq_regs();
 361         int duration;
 362         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 363
 364         if (!watchdog_enabled)
 365                 return HRTIMER_NORESTART;
 366
 367         /* kick the hardlockup detector */
 368         watchdog_interrupt_count();
 369
 370         /* kick the softlockup detector */
 371         if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 372                 reinit_completion(this_cpu_ptr(&softlockup_completion));
 373                 stop_one_cpu_nowait(smp_processor_id(),
 374                                 softlockup_fn, NULL,
 375                                 this_cpu_ptr(&softlockup_stop_work));
 376         }
 377
 378         /* .. and repeat */
 379         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 380
 381         if (touch_ts == SOFTLOCKUP_RESET) {
 382                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 383                         /*
 384                          * If the time stamp was touched atomically
 385                          * make sure the scheduler tick is up to date.
 386                          */
 387                         __this_cpu_write(softlockup_touch_sync, false);
 388                         sched_clock_tick();
 389                 }
 390
 391                 /* Clear the guest paused flag on watchdog reset */
 392                 kvm_check_and_clear_guest_paused();
 393                 __touch_watchdog();
 394                 return HRTIMER_RESTART;
 395         }
 396
 397         /* check for a softlockup
 398          * This is done by making sure a high priority task is
 399          * being scheduled.  The task touches the watchdog to
 400          * indicate it is getting cpu time.  If it hasn't then
 401          * this is a good indication some task is hogging the cpu
 402          */
 403         duration = is_softlockup(touch_ts);
 404         if (unlikely(duration)) {
 405                 /*
 406                  * If a virtual machine is stopped by the host it can look to
 407                  * the watchdog like a soft lockup, check to see if the host
 408                  * stopped the vm before we issue the warning
 409                  */
 410                 if (kvm_check_and_clear_guest_paused())
 411                         return HRTIMER_RESTART;
 412
 413                 /* only warn once */
 414                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 415                         /*
 416                          * When multiple processes are causing softlockups the
 417                          * softlockup detector only warns on the first one
 418                          * because the code relies on a full quiet cycle to
 419                          * re-arm.  The second process prevents the quiet cycle
 420                          * and never gets reported.  Use task pointers to detect
 421                          * this.
 422                          */
 423                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 424                             current) {
 425                                 __this_cpu_write(soft_watchdog_warn, false);
 426                                 __touch_watchdog();
 427                         }
 428                         return HRTIMER_RESTART;
 429                 }
 430
 431                 if (softlockup_all_cpu_backtrace) {
 432                         /* Prevent multiple soft-lockup reports if one cpu is already
 433                          * engaged in dumping cpu back traces
 434                          */
 435                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 436                                 /* Someone else will report us. Let's give up */
 437                                 __this_cpu_write(soft_watchdog_warn, true);
 438                                 return HRTIMER_RESTART;
 439                         }
 440                 }
 441
 442                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 443                         smp_processor_id(), duration,
 444                         current->comm, task_pid_nr(current));
 445                 __this_cpu_write(softlockup_task_ptr_saved, current);
 446                 print_modules();
 447                 print_irqtrace_events(current);
 448                 if (regs)
 449                         show_regs(regs);
 450                 else
 451                         dump_stack();
 452
 453                 if (softlockup_all_cpu_backtrace) {
 454                         /* Avoid generating two back traces for current
 455                          * given that one is already made above
 456                          */
 457                         trigger_allbutself_cpu_backtrace();
 458
 459                         clear_bit(0, &soft_lockup_nmi_warn);
 460                         /* Barrier to sync with other cpus */
 461                         smp_mb__after_atomic();
 462                 }
 463
 464                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 465                 if (softlockup_panic)
 466                         panic("softlockup: hung tasks");
 467                 __this_cpu_write(soft_watchdog_warn, true);
 468         } else
 469                 __this_cpu_write(soft_watchdog_warn, false);
 470
 471         return HRTIMER_RESTART;
 472 }
 473
 474 static void watchdog_enable(unsigned int cpu)
 475 {
 476         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 477         struct completion *done = this_cpu_ptr(&softlockup_completion);
 478
 479         WARN_ON_ONCE(cpu != smp_processor_id());
 480
 481         init_completion(done);
 482         complete(done);
 483
 484         /*
 485          * Start the timer first to prevent the NMI watchdog triggering
 486          * before the timer has a chance to fire.
 487          */
 488         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 489         hrtimer->function = watchdog_timer_fn;
 490         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 491                       HRTIMER_MODE_REL_PINNED);
 492
 493         /* Initialize timestamp */
 494         __touch_watchdog();
 495         /* Enable the perf event */
 496         if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
 497                 watchdog_nmi_enable(cpu);
 498 }
 499
 500 static void watchdog_disable(unsigned int cpu)
 501 {
 502         struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 503
 504         WARN_ON_ONCE(cpu != smp_processor_id());
 505
 506         /*
 507          * Disable the perf event first. That prevents that a large delay
 508          * between disabling the timer and disabling the perf event causes
 509          * the perf NMI to detect a false positive.
 510          */
 511         watchdog_nmi_disable(cpu);
 512         hrtimer_cancel(hrtimer);
 513         wait_for_completion(this_cpu_ptr(&softlockup_completion));
 514 }
 515
 516 static int softlockup_stop_fn(void *data)
 517 {
 518         watchdog_disable(smp_processor_id());
 519         return 0;
 520 }
 521
 522 static void softlockup_stop_all(void)
 523 {
 524         int cpu;
 525
 526         if (!softlockup_initialized)
 527                 return;
 528
 529         for_each_cpu(cpu, &watchdog_allowed_mask)
 530                 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 531
 532         cpumask_clear(&watchdog_allowed_mask);
 533 }
 534
 535 static int softlockup_start_fn(void *data)
 536 {
 537         watchdog_enable(smp_processor_id());
 538         return 0;
 539 }
 540
 541 static void softlockup_start_all(void)
 542 {
 543         int cpu;
 544
 545         cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 546         for_each_cpu(cpu, &watchdog_allowed_mask)
 547                 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 548 }
 549
 550 int lockup_detector_online_cpu(unsigned int cpu)
 551 {
 552         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 553                 watchdog_enable(cpu);
 554         return 0;
 555 }
 556
 557 int lockup_detector_offline_cpu(unsigned int cpu)
 558 {
 559         if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 560                 watchdog_disable(cpu);
 561         return 0;
 562 }
 563
 564 static void __lockup_detector_reconfigure(void)
 565 {
 566         cpus_read_lock();
 567         watchdog_nmi_stop();
 568
 569         softlockup_stop_all();
 570         set_sample_period();
 571         lockup_detector_update_enable();
 572         if (watchdog_enabled && watchdog_thresh)
 573                 softlockup_start_all();
 574
 575         watchdog_nmi_start();
 576         cpus_read_unlock();
 577         /*
 578          * Must be called outside the cpus locked section to prevent
 579          * recursive locking in the perf code.
 580          */
 581         __lockup_detector_cleanup();
 582 }
 583
 584 void lockup_detector_reconfigure(void)
 585 {
 586         mutex_lock(&watchdog_mutex);
 587         __lockup_detector_reconfigure();
 588         mutex_unlock(&watchdog_mutex);
 589 }
 590
 591 /*
 592  * Create the watchdog thread infrastructure and configure the detector(s).
 593  *
 594  * The threads are not unparked as watchdog_allowed_mask is empty.  When
 595  * the threads are sucessfully initialized, take the proper locks and
 596  * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
 597  */
 598 static __init void lockup_detector_setup(void)
 599 {
 600         /*
 601          * If sysctl is off and watchdog got disabled on the command line,
 602          * nothing to do here.
 603          */
 604         lockup_detector_update_enable();
 605
 606         if (!IS_ENABLED(CONFIG_SYSCTL) &&
 607             !(watchdog_enabled && watchdog_thresh))
 608                 return;
 609
 610         mutex_lock(&watchdog_mutex);
 611         __lockup_detector_reconfigure();
 612         softlockup_initialized = true;
 613         mutex_unlock(&watchdog_mutex);
 614 }
 615
 616 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
 617 static void __lockup_detector_reconfigure(void)
 618 {
 619         cpus_read_lock();
 620         watchdog_nmi_stop();
 621         lockup_detector_update_enable();
 622         watchdog_nmi_start();
 623         cpus_read_unlock();
 624 }
 625 void lockup_detector_reconfigure(void)
 626 {
 627         __lockup_detector_reconfigure();
 628 }
 629 static inline void lockup_detector_setup(void)
 630 {
 631         __lockup_detector_reconfigure();
 632 }
 633 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 634
 635 static void __lockup_detector_cleanup(void)
 636 {
 637         lockdep_assert_held(&watchdog_mutex);
 638         hardlockup_detector_perf_cleanup();
 639 }
 640
 641 /**
 642  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 643  *
 644  * Caller must not hold the cpu hotplug rwsem.
 645  */
 646 void lockup_detector_cleanup(void)
 647 {
 648         mutex_lock(&watchdog_mutex);
 649         __lockup_detector_cleanup();
 650         mutex_unlock(&watchdog_mutex);
 651 }
 652
 653 /**
 654  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 655  *
 656  * Special interface for parisc. It prevents lockup detector warnings from
 657  * the default pm_poweroff() function which busy loops forever.
 658  */
 659 void lockup_detector_soft_poweroff(void)
 660 {
 661         watchdog_enabled = 0;
 662 }
 663
 664 #ifdef CONFIG_SYSCTL
 665
 666 /* Propagate any changes to the watchdog threads */
 667 static void proc_watchdog_update(void)
 668 {
 669         /* Remove impossible cpus to keep sysctl output clean. */
 670         cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 671         __lockup_detector_reconfigure();
 672 }
 673
 674 /*
 675  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 676  *
 677  * caller             | table->data points to      | 'which'
 678  * -------------------|----------------------------|--------------------------
 679  * proc_watchdog      | watchdog_user_enabled      | NMI_WATCHDOG_ENABLED |
 680  *                    |                            | SOFT_WATCHDOG_ENABLED
 681  * -------------------|----------------------------|--------------------------
 682  * proc_nmi_watchdog  | nmi_watchdog_user_enabled  | NMI_WATCHDOG_ENABLED
 683  * -------------------|----------------------------|--------------------------
 684  * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
 685  */
 686 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 687                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 688 {
 689         int err, old, *param = table->data;
 690
 691         mutex_lock(&watchdog_mutex);
 692
 693         if (!write) {
 694                 /*
 695                  * On read synchronize the userspace interface. This is a
 696                  * racy snapshot.
 697                  */
 698                 *param = (watchdog_enabled & which) != 0;
 699                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 700         } else {
 701                 old = READ_ONCE(*param);
 702                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 703                 if (!err && old != READ_ONCE(*param))
 704                         proc_watchdog_update();
 705         }
 706         mutex_unlock(&watchdog_mutex);
 707         return err;
 708 }
 709
 710 /*
 711  * /proc/sys/kernel/watchdog
 712  */
 713 int proc_watchdog(struct ctl_table *table, int write,
 714                   void __user *buffer, size_t *lenp, loff_t *ppos)
 715 {
 716         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 717                                     table, write, buffer, lenp, ppos);
 718 }
 719
 720 /*
 721  * /proc/sys/kernel/nmi_watchdog
 722  */
 723 int proc_nmi_watchdog(struct ctl_table *table, int write,
 724                       void __user *buffer, size_t *lenp, loff_t *ppos)
 725 {
 726         if (!nmi_watchdog_available && write)
 727                 return -ENOTSUPP;
 728         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 729                                     table, write, buffer, lenp, ppos);
 730 }
 731
 732 /*
 733  * /proc/sys/kernel/soft_watchdog
 734  */
 735 int proc_soft_watchdog(struct ctl_table *table, int write,
 736                         void __user *buffer, size_t *lenp, loff_t *ppos)
 737 {
 738         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 739                                     table, write, buffer, lenp, ppos);
 740 }
 741
 742 /*
 743  * /proc/sys/kernel/watchdog_thresh
 744  */
 745 int proc_watchdog_thresh(struct ctl_table *table, int write,
 746                          void __user *buffer, size_t *lenp, loff_t *ppos)
 747 {
 748         int err, old;
 749
 750         mutex_lock(&watchdog_mutex);
 751
 752         old = READ_ONCE(watchdog_thresh);
 753         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 754
 755         if (!err && write && old != READ_ONCE(watchdog_thresh))
 756                 proc_watchdog_update();
 757
 758         mutex_unlock(&watchdog_mutex);
 759         return err;
 760 }
 761
 762 /*
 763  * The cpumask is the mask of possible cpus that the watchdog can run
 764  * on, not the mask of cpus it is actually running on.  This allows the
 765  * user to specify a mask that will include cpus that have not yet
 766  * been brought online, if desired.
 767  */
 768 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 769                           void __user *buffer, size_t *lenp, loff_t *ppos)
 770 {
 771         int err;
 772
 773         mutex_lock(&watchdog_mutex);
 774
 775         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 776         if (!err && write)
 777                 proc_watchdog_update();
 778
 779         mutex_unlock(&watchdog_mutex);
 780         return err;
 781 }
 782 #endif /* CONFIG_SYSCTL */
 783
 784 void __init lockup_detector_init(void)
 785 {
 786         if (tick_nohz_full_enabled())
 787                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 788
 789         cpumask_copy(&watchdog_cpumask,
 790                      housekeeping_cpumask(HK_FLAG_TIMER));
 791
 792         if (!watchdog_nmi_probe())
 793                 nmi_watchdog_available = true;
 794         lockup_detector_setup();
 795 }