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