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