kernel/watchdog.c

   1 /*
   2  * Detect hard and soft lockups on a system
   3  *
   4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   5  *
   6  * Note: Most of this code is borrowed heavily from the original softlockup
   7  * detector, so thanks to Ingo for the initial implementation.
   8  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
   9  * to those contributors as well.
  10  */
  11
  12 #define pr_fmt(fmt) "watchdog: " fmt
  13
  14 #include <linux/mm.h>
  15 #include <linux/cpu.h>
  16 #include <linux/nmi.h>
  17 #include <linux/init.h>
  18 #include <linux/module.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/smpboot.h>
  21 #include <linux/sched/rt.h>
  22 #include <uapi/linux/sched/types.h>
  23 #include <linux/tick.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/sched/clock.h>
  26 #include <linux/sched/debug.h>
  27
  28 #include <asm/irq_regs.h>
  29 #include <linux/kvm_para.h>
  30 #include <linux/kthread.h>
  31
  32 /* Watchdog configuration */
  33 static DEFINE_MUTEX(watchdog_proc_mutex);
  34
  35 int __read_mostly nmi_watchdog_enabled;
  36
  37 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
  38 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED |
  39                                                 NMI_WATCHDOG_ENABLED;
  40 #else
  41 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  42 #endif
  43
  44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  45 /* boot commands */
  46 /*
  47  * Should we panic when a soft-lockup or hard-lockup occurs:
  48  */
  49 unsigned int __read_mostly hardlockup_panic =
  50                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  51 /*
  52  * We may not want to enable hard lockup detection by default in all cases,
  53  * for example when running the kernel as a guest on a hypervisor. In these
  54  * cases this function can be called to disable hard lockup detection. This
  55  * function should only be executed once by the boot processor before the
  56  * kernel command line parameters are parsed, because otherwise it is not
  57  * possible to override this in hardlockup_panic_setup().
  58  */
  59 void hardlockup_detector_disable(void)
  60 {
  61         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  62 }
  63
  64 static int __init hardlockup_panic_setup(char *str)
  65 {
  66         if (!strncmp(str, "panic", 5))
  67                 hardlockup_panic = 1;
  68         else if (!strncmp(str, "nopanic", 7))
  69                 hardlockup_panic = 0;
  70         else if (!strncmp(str, "0", 1))
  71                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  72         else if (!strncmp(str, "1", 1))
  73                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
  74         return 1;
  75 }
  76 __setup("nmi_watchdog=", hardlockup_panic_setup);
  77
  78 #endif
  79
  80 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
  81 int __read_mostly soft_watchdog_enabled;
  82 #endif
  83
  84 int __read_mostly watchdog_user_enabled;
  85 int __read_mostly watchdog_thresh = 10;
  86
  87 #ifdef CONFIG_SMP
  88 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  89 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  90 #endif
  91 struct cpumask watchdog_cpumask __read_mostly;
  92 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  93
  94 /*
  95  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  96  * are registered/started and is set to 0 when the watchdog threads are
  97  * unregistered/stopped, so it is an indicator whether the threads exist.
  98  */
  99 static int __read_mostly watchdog_running;
 100
 101 /*
 102  * These functions can be overridden if an architecture implements its
 103  * own hardlockup detector.
 104  *
 105  * watchdog_nmi_enable/disable can be implemented to start and stop when
 106  * softlockup watchdog threads start and stop. The arch must select the
 107  * SOFTLOCKUP_DETECTOR Kconfig.
 108  */
 109 int __weak watchdog_nmi_enable(unsigned int cpu)
 110 {
 111         return 0;
 112 }
 113 void __weak watchdog_nmi_disable(unsigned int cpu)
 114 {
 115 }
 116
 117 /*
 118  * watchdog_nmi_reconfigure can be implemented to be notified after any
 119  * watchdog configuration change. The arch hardlockup watchdog should
 120  * respond to the following variables:
 121  * - nmi_watchdog_enabled
 122  * - watchdog_thresh
 123  * - watchdog_cpumask
 124  * - sysctl_hardlockup_all_cpu_backtrace
 125  * - hardlockup_panic
 126  */
 127 void __weak watchdog_nmi_reconfigure(void)
 128 {
 129 }
 130
 131
 132 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 133
 134 /* Helper for online, unparked cpus. */
 135 #define for_each_watchdog_cpu(cpu) \
 136         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
 137
 138 atomic_t watchdog_park_in_progress = ATOMIC_INIT(0);
 139
 140 static u64 __read_mostly sample_period;
 141
 142 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 143 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
 144 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 145 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 146 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 147 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 148 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 149 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 150 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 151 static unsigned long soft_lockup_nmi_warn;
 152
 153 unsigned int __read_mostly softlockup_panic =
 154                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 155
 156 static int __init softlockup_panic_setup(char *str)
 157 {
 158         softlockup_panic = simple_strtoul(str, NULL, 0);
 159
 160         return 1;
 161 }
 162 __setup("softlockup_panic=", softlockup_panic_setup);
 163
 164 static int __init nowatchdog_setup(char *str)
 165 {
 166         watchdog_enabled = 0;
 167         return 1;
 168 }
 169 __setup("nowatchdog", nowatchdog_setup);
 170
 171 static int __init nosoftlockup_setup(char *str)
 172 {
 173         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 174         return 1;
 175 }
 176 __setup("nosoftlockup", nosoftlockup_setup);
 177
 178 #ifdef CONFIG_SMP
 179 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 180 {
 181         sysctl_softlockup_all_cpu_backtrace =
 182                 !!simple_strtol(str, NULL, 0);
 183         return 1;
 184 }
 185 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 186 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 187 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
 188 {
 189         sysctl_hardlockup_all_cpu_backtrace =
 190                 !!simple_strtol(str, NULL, 0);
 191         return 1;
 192 }
 193 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
 194 #endif
 195 #endif
 196
 197 /*
 198  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 199  * lockups can have false positives under extreme conditions. So we generally
 200  * want a higher threshold for soft lockups than for hard lockups. So we couple
 201  * the thresholds with a factor: we make the soft threshold twice the amount of
 202  * time the hard threshold is.
 203  */
 204 static int get_softlockup_thresh(void)
 205 {
 206         return watchdog_thresh * 2;
 207 }
 208
 209 /*
 210  * Returns seconds, approximately.  We don't need nanosecond
 211  * resolution, and we don't need to waste time with a big divide when
 212  * 2^30ns == 1.074s.
 213  */
 214 static unsigned long get_timestamp(void)
 215 {
 216         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 217 }
 218
 219 static void set_sample_period(void)
 220 {
 221         /*
 222          * convert watchdog_thresh from seconds to ns
 223          * the divide by 5 is to give hrtimer several chances (two
 224          * or three with the current relation between the soft
 225          * and hard thresholds) to increment before the
 226          * hardlockup detector generates a warning
 227          */
 228         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 229         watchdog_update_hrtimer_threshold(sample_period);
 230 }
 231
 232 /* Commands for resetting the watchdog */
 233 static void __touch_watchdog(void)
 234 {
 235         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 236 }
 237
 238 /**
 239  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 240  *
 241  * Call when the scheduler may have stalled for legitimate reasons
 242  * preventing the watchdog task from executing - e.g. the scheduler
 243  * entering idle state.  This should only be used for scheduler events.
 244  * Use touch_softlockup_watchdog() for everything else.
 245  */
 246 void touch_softlockup_watchdog_sched(void)
 247 {
 248         /*
 249          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 250          * gets zeroed here, so use the raw_ operation.
 251          */
 252         raw_cpu_write(watchdog_touch_ts, 0);
 253 }
 254
 255 void touch_softlockup_watchdog(void)
 256 {
 257         touch_softlockup_watchdog_sched();
 258         wq_watchdog_touch(raw_smp_processor_id());
 259 }
 260 EXPORT_SYMBOL(touch_softlockup_watchdog);
 261
 262 void touch_all_softlockup_watchdogs(void)
 263 {
 264         int cpu;
 265
 266         /*
 267          * this is done lockless
 268          * do we care if a 0 races with a timestamp?
 269          * all it means is the softlock check starts one cycle later
 270          */
 271         for_each_watchdog_cpu(cpu)
 272                 per_cpu(watchdog_touch_ts, cpu) = 0;
 273         wq_watchdog_touch(-1);
 274 }
 275
 276 void touch_softlockup_watchdog_sync(void)
 277 {
 278         __this_cpu_write(softlockup_touch_sync, true);
 279         __this_cpu_write(watchdog_touch_ts, 0);
 280 }
 281
 282 static int is_softlockup(unsigned long touch_ts)
 283 {
 284         unsigned long now = get_timestamp();
 285
 286         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 287                 /* Warn about unreasonable delays. */
 288                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 289                         return now - touch_ts;
 290         }
 291         return 0;
 292 }
 293
 294 /* watchdog detector functions */
 295 bool is_hardlockup(void)
 296 {
 297         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 298
 299         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 300                 return true;
 301
 302         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 303         return false;
 304 }
 305
 306 static void watchdog_interrupt_count(void)
 307 {
 308         __this_cpu_inc(hrtimer_interrupts);
 309 }
 310
 311 static int watchdog_enable_all_cpus(void);
 312 static void watchdog_disable_all_cpus(void);
 313
 314 /* watchdog kicker functions */
 315 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 316 {
 317         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 318         struct pt_regs *regs = get_irq_regs();
 319         int duration;
 320         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 321
 322         if (!watchdog_enabled ||
 323             atomic_read(&watchdog_park_in_progress) != 0)
 324                 return HRTIMER_NORESTART;
 325
 326         /* kick the hardlockup detector */
 327         watchdog_interrupt_count();
 328
 329         /* kick the softlockup detector */
 330         wake_up_process(__this_cpu_read(softlockup_watchdog));
 331
 332         /* .. and repeat */
 333         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 334
 335         if (touch_ts == 0) {
 336                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 337                         /*
 338                          * If the time stamp was touched atomically
 339                          * make sure the scheduler tick is up to date.
 340                          */
 341                         __this_cpu_write(softlockup_touch_sync, false);
 342                         sched_clock_tick();
 343                 }
 344
 345                 /* Clear the guest paused flag on watchdog reset */
 346                 kvm_check_and_clear_guest_paused();
 347                 __touch_watchdog();
 348                 return HRTIMER_RESTART;
 349         }
 350
 351         /* check for a softlockup
 352          * This is done by making sure a high priority task is
 353          * being scheduled.  The task touches the watchdog to
 354          * indicate it is getting cpu time.  If it hasn't then
 355          * this is a good indication some task is hogging the cpu
 356          */
 357         duration = is_softlockup(touch_ts);
 358         if (unlikely(duration)) {
 359                 /*
 360                  * If a virtual machine is stopped by the host it can look to
 361                  * the watchdog like a soft lockup, check to see if the host
 362                  * stopped the vm before we issue the warning
 363                  */
 364                 if (kvm_check_and_clear_guest_paused())
 365                         return HRTIMER_RESTART;
 366
 367                 /* only warn once */
 368                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 369                         /*
 370                          * When multiple processes are causing softlockups the
 371                          * softlockup detector only warns on the first one
 372                          * because the code relies on a full quiet cycle to
 373                          * re-arm.  The second process prevents the quiet cycle
 374                          * and never gets reported.  Use task pointers to detect
 375                          * this.
 376                          */
 377                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 378                             current) {
 379                                 __this_cpu_write(soft_watchdog_warn, false);
 380                                 __touch_watchdog();
 381                         }
 382                         return HRTIMER_RESTART;
 383                 }
 384
 385                 if (softlockup_all_cpu_backtrace) {
 386                         /* Prevent multiple soft-lockup reports if one cpu is already
 387                          * engaged in dumping cpu back traces
 388                          */
 389                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 390                                 /* Someone else will report us. Let's give up */
 391                                 __this_cpu_write(soft_watchdog_warn, true);
 392                                 return HRTIMER_RESTART;
 393                         }
 394                 }
 395
 396                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 397                         smp_processor_id(), duration,
 398                         current->comm, task_pid_nr(current));
 399                 __this_cpu_write(softlockup_task_ptr_saved, current);
 400                 print_modules();
 401                 print_irqtrace_events(current);
 402                 if (regs)
 403                         show_regs(regs);
 404                 else
 405                         dump_stack();
 406
 407                 if (softlockup_all_cpu_backtrace) {
 408                         /* Avoid generating two back traces for current
 409                          * given that one is already made above
 410                          */
 411                         trigger_allbutself_cpu_backtrace();
 412
 413                         clear_bit(0, &soft_lockup_nmi_warn);
 414                         /* Barrier to sync with other cpus */
 415                         smp_mb__after_atomic();
 416                 }
 417
 418                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 419                 if (softlockup_panic)
 420                         panic("softlockup: hung tasks");
 421                 __this_cpu_write(soft_watchdog_warn, true);
 422         } else
 423                 __this_cpu_write(soft_watchdog_warn, false);
 424
 425         return HRTIMER_RESTART;
 426 }
 427
 428 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 429 {
 430         struct sched_param param = { .sched_priority = prio };
 431
 432         sched_setscheduler(current, policy, &param);
 433 }
 434
 435 static void watchdog_enable(unsigned int cpu)
 436 {
 437         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 438
 439         /* kick off the timer for the hardlockup detector */
 440         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 441         hrtimer->function = watchdog_timer_fn;
 442
 443         /* Enable the perf event */
 444         watchdog_nmi_enable(cpu);
 445
 446         /* done here because hrtimer_start can only pin to smp_processor_id() */
 447         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 448                       HRTIMER_MODE_REL_PINNED);
 449
 450         /* initialize timestamp */
 451         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 452         __touch_watchdog();
 453 }
 454
 455 static void watchdog_disable(unsigned int cpu)
 456 {
 457         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 458
 459         watchdog_set_prio(SCHED_NORMAL, 0);
 460         hrtimer_cancel(hrtimer);
 461         /* disable the perf event */
 462         watchdog_nmi_disable(cpu);
 463 }
 464
 465 static void watchdog_cleanup(unsigned int cpu, bool online)
 466 {
 467         watchdog_disable(cpu);
 468 }
 469
 470 static int watchdog_should_run(unsigned int cpu)
 471 {
 472         return __this_cpu_read(hrtimer_interrupts) !=
 473                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 474 }
 475
 476 /*
 477  * The watchdog thread function - touches the timestamp.
 478  *
 479  * It only runs once every sample_period seconds (4 seconds by
 480  * default) to reset the softlockup timestamp. If this gets delayed
 481  * for more than 2*watchdog_thresh seconds then the debug-printout
 482  * triggers in watchdog_timer_fn().
 483  */
 484 static void watchdog(unsigned int cpu)
 485 {
 486         __this_cpu_write(soft_lockup_hrtimer_cnt,
 487                          __this_cpu_read(hrtimer_interrupts));
 488         __touch_watchdog();
 489
 490         /*
 491          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 492          * failure path. Check for failures that can occur asynchronously -
 493          * for example, when CPUs are on-lined - and shut down the hardware
 494          * perf event on each CPU accordingly.
 495          *
 496          * The only non-obvious place this bit can be cleared is through
 497          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 498          * pr_info here would be too noisy as it would result in a message
 499          * every few seconds if the hardlockup was disabled but the softlockup
 500          * enabled.
 501          */
 502         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 503                 watchdog_nmi_disable(cpu);
 504 }
 505
 506 static struct smp_hotplug_thread watchdog_threads = {
 507         .store                  = &softlockup_watchdog,
 508         .thread_should_run      = watchdog_should_run,
 509         .thread_fn              = watchdog,
 510         .thread_comm            = "watchdog/%u",
 511         .setup                  = watchdog_enable,
 512         .cleanup                = watchdog_cleanup,
 513         .park                   = watchdog_disable,
 514         .unpark                 = watchdog_enable,
 515 };
 516
 517 /*
 518  * park all watchdog threads that are specified in 'watchdog_cpumask'
 519  *
 520  * This function returns an error if kthread_park() of a watchdog thread
 521  * fails. In this situation, the watchdog threads of some CPUs can already
 522  * be parked and the watchdog threads of other CPUs can still be runnable.
 523  * Callers are expected to handle this special condition as appropriate in
 524  * their context.
 525  *
 526  * This function may only be called in a context that is protected against
 527  * races with CPU hotplug - for example, via get_online_cpus().
 528  */
 529 static int watchdog_park_threads(void)
 530 {
 531         int cpu, ret = 0;
 532
 533         atomic_set(&watchdog_park_in_progress, 1);
 534
 535         for_each_watchdog_cpu(cpu) {
 536                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 537                 if (ret)
 538                         break;
 539         }
 540
 541         atomic_set(&watchdog_park_in_progress, 0);
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 548  *
 549  * This function may only be called in a context that is protected against
 550  * races with CPU hotplug - for example, via get_online_cpus().
 551  */
 552 static void watchdog_unpark_threads(void)
 553 {
 554         int cpu;
 555
 556         for_each_watchdog_cpu(cpu)
 557                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 558 }
 559
 560 static int update_watchdog_all_cpus(void)
 561 {
 562         int ret;
 563
 564         ret = watchdog_park_threads();
 565         if (ret)
 566                 return ret;
 567
 568         watchdog_unpark_threads();
 569
 570         return 0;
 571 }
 572
 573 static int watchdog_enable_all_cpus(void)
 574 {
 575         int err = 0;
 576
 577         if (!watchdog_running) {
 578                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 579                                                              &watchdog_cpumask);
 580                 if (err)
 581                         pr_err("Failed to create watchdog threads, disabled\n");
 582                 else
 583                         watchdog_running = 1;
 584         } else {
 585                 /*
 586                  * Enable/disable the lockup detectors or
 587                  * change the sample period 'on the fly'.
 588                  */
 589                 err = update_watchdog_all_cpus();
 590
 591                 if (err) {
 592                         watchdog_disable_all_cpus();
 593                         pr_err("Failed to update lockup detectors, disabled\n");
 594                 }
 595         }
 596
 597         if (err)
 598                 watchdog_enabled = 0;
 599
 600         return err;
 601 }
 602
 603 static void watchdog_disable_all_cpus(void)
 604 {
 605         if (watchdog_running) {
 606                 watchdog_running = 0;
 607                 smpboot_unregister_percpu_thread(&watchdog_threads);
 608         }
 609 }
 610
 611 #ifdef CONFIG_SYSCTL
 612 static int watchdog_update_cpus(void)
 613 {
 614         return smpboot_update_cpumask_percpu_thread(
 615                     &watchdog_threads, &watchdog_cpumask);
 616 }
 617 #endif
 618
 619 #else /* SOFTLOCKUP */
 620 static int watchdog_park_threads(void)
 621 {
 622         return 0;
 623 }
 624
 625 static void watchdog_unpark_threads(void)
 626 {
 627 }
 628
 629 static int watchdog_enable_all_cpus(void)
 630 {
 631         return 0;
 632 }
 633
 634 static void watchdog_disable_all_cpus(void)
 635 {
 636 }
 637
 638 #ifdef CONFIG_SYSCTL
 639 static int watchdog_update_cpus(void)
 640 {
 641         return 0;
 642 }
 643 #endif
 644
 645 static void set_sample_period(void)
 646 {
 647 }
 648 #endif /* SOFTLOCKUP */
 649
 650 /**
 651  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 652  *
 653  * Special interface for parisc. It prevents lockup detector warnings from
 654  * the default pm_poweroff() function which busy loops forever.
 655  */
 656 void lockup_detector_soft_poweroff(void)
 657 {
 658         watchdog_enabled = 0;
 659 }
 660
 661 #ifdef CONFIG_SYSCTL
 662
 663 /*
 664  * Update the run state of the lockup detectors.
 665  */
 666 static int proc_watchdog_update(void)
 667 {
 668         int err = 0;
 669
 670         /*
 671          * Watchdog threads won't be started if they are already active.
 672          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 673          * care of this. If those threads are already active, the sample
 674          * period will be updated and the lockup detectors will be enabled
 675          * or disabled 'on the fly'.
 676          */
 677         if (watchdog_enabled && watchdog_thresh)
 678                 err = watchdog_enable_all_cpus();
 679         else
 680                 watchdog_disable_all_cpus();
 681
 682         watchdog_nmi_reconfigure();
 683
 684         return err;
 685
 686 }
 687
 688 /*
 689  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 690  *
 691  * caller             | table->data points to | 'which' contains the flag(s)
 692  * -------------------|-----------------------|-----------------------------
 693  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 694  *                    |                       | with SOFT_WATCHDOG_ENABLED
 695  * -------------------|-----------------------|-----------------------------
 696  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 697  * -------------------|-----------------------|-----------------------------
 698  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 699  */
 700 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 701                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 702 {
 703         int err, old, new;
 704         int *watchdog_param = (int *)table->data;
 705
 706         get_online_cpus();
 707         mutex_lock(&watchdog_proc_mutex);
 708
 709         /*
 710          * If the parameter is being read return the state of the corresponding
 711          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 712          * run state of the lockup detectors.
 713          */
 714         if (!write) {
 715                 *watchdog_param = (watchdog_enabled & which) != 0;
 716                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 717         } else {
 718                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 719                 if (err)
 720                         goto out;
 721
 722                 /*
 723                  * There is a race window between fetching the current value
 724                  * from 'watchdog_enabled' and storing the new value. During
 725                  * this race window, watchdog_nmi_enable() can sneak in and
 726                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 727                  * The 'cmpxchg' detects this race and the loop retries.
 728                  */
 729                 do {
 730                         old = watchdog_enabled;
 731                         /*
 732                          * If the parameter value is not zero set the
 733                          * corresponding bit(s), else clear it(them).
 734                          */
 735                         if (*watchdog_param)
 736                                 new = old | which;
 737                         else
 738                                 new = old & ~which;
 739                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 740
 741                 /*
 742                  * Update the run state of the lockup detectors. There is _no_
 743                  * need to check the value returned by proc_watchdog_update()
 744                  * and to restore the previous value of 'watchdog_enabled' as
 745                  * both lockup detectors are disabled if proc_watchdog_update()
 746                  * returns an error.
 747                  */
 748                 if (old == new)
 749                         goto out;
 750
 751                 err = proc_watchdog_update();
 752         }
 753 out:
 754         mutex_unlock(&watchdog_proc_mutex);
 755         put_online_cpus();
 756         return err;
 757 }
 758
 759 /*
 760  * /proc/sys/kernel/watchdog
 761  */
 762 int proc_watchdog(struct ctl_table *table, int write,
 763                   void __user *buffer, size_t *lenp, loff_t *ppos)
 764 {
 765         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 766                                     table, write, buffer, lenp, ppos);
 767 }
 768
 769 /*
 770  * /proc/sys/kernel/nmi_watchdog
 771  */
 772 int proc_nmi_watchdog(struct ctl_table *table, int write,
 773                       void __user *buffer, size_t *lenp, loff_t *ppos)
 774 {
 775         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 776                                     table, write, buffer, lenp, ppos);
 777 }
 778
 779 /*
 780  * /proc/sys/kernel/soft_watchdog
 781  */
 782 int proc_soft_watchdog(struct ctl_table *table, int write,
 783                         void __user *buffer, size_t *lenp, loff_t *ppos)
 784 {
 785         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 786                                     table, write, buffer, lenp, ppos);
 787 }
 788
 789 /*
 790  * /proc/sys/kernel/watchdog_thresh
 791  */
 792 int proc_watchdog_thresh(struct ctl_table *table, int write,
 793                          void __user *buffer, size_t *lenp, loff_t *ppos)
 794 {
 795         int err, old, new;
 796
 797         get_online_cpus();
 798         mutex_lock(&watchdog_proc_mutex);
 799
 800         old = ACCESS_ONCE(watchdog_thresh);
 801         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 802
 803         if (err || !write)
 804                 goto out;
 805
 806         /*
 807          * Update the sample period. Restore on failure.
 808          */
 809         new = ACCESS_ONCE(watchdog_thresh);
 810         if (old == new)
 811                 goto out;
 812
 813         set_sample_period();
 814         err = proc_watchdog_update();
 815         if (err) {
 816                 watchdog_thresh = old;
 817                 set_sample_period();
 818         }
 819 out:
 820         mutex_unlock(&watchdog_proc_mutex);
 821         put_online_cpus();
 822         return err;
 823 }
 824
 825 /*
 826  * The cpumask is the mask of possible cpus that the watchdog can run
 827  * on, not the mask of cpus it is actually running on.  This allows the
 828  * user to specify a mask that will include cpus that have not yet
 829  * been brought online, if desired.
 830  */
 831 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 832                           void __user *buffer, size_t *lenp, loff_t *ppos)
 833 {
 834         int err;
 835
 836         get_online_cpus();
 837         mutex_lock(&watchdog_proc_mutex);
 838
 839         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 840         if (!err && write) {
 841                 /* Remove impossible cpus to keep sysctl output cleaner. */
 842                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
 843                             cpu_possible_mask);
 844
 845                 if (watchdog_running) {
 846                         /*
 847                          * Failure would be due to being unable to allocate
 848                          * a temporary cpumask, so we are likely not in a
 849                          * position to do much else to make things better.
 850                          */
 851                         if (watchdog_update_cpus() != 0)
 852                                 pr_err("cpumask update failed\n");
 853                 }
 854
 855                 watchdog_nmi_reconfigure();
 856         }
 857
 858         mutex_unlock(&watchdog_proc_mutex);
 859         put_online_cpus();
 860         return err;
 861 }
 862
 863 #endif /* CONFIG_SYSCTL */
 864
 865 void __init lockup_detector_init(void)
 866 {
 867         set_sample_period();
 868
 869 #ifdef CONFIG_NO_HZ_FULL
 870         if (tick_nohz_full_enabled()) {
 871                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 872                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
 873         } else
 874                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 875 #else
 876         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 877 #endif
 878
 879         if (watchdog_enabled)
 880                 watchdog_enable_all_cpus();
 881 }