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) "NMI 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 <linux/tick.h>
  23
  24 #include <asm/irq_regs.h>
  25 #include <linux/kvm_para.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/kthread.h>
  28
  29 /*
  30  * The run state of the lockup detectors is controlled by the content of the
  31  * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
  32  * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
  33  *
  34  * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
  35  * are variables that are only used as an 'interface' between the parameters
  36  * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
  37  * 'watchdog_thresh' variable is handled differently because its value is not
  38  * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
  39  * is equal zero.
  40  */
  41 #define NMI_WATCHDOG_ENABLED_BIT   0
  42 #define SOFT_WATCHDOG_ENABLED_BIT  1
  43 #define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
  44 #define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
  45
  46 static DEFINE_MUTEX(watchdog_proc_mutex);
  47
  48 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  49 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
  50 #else
  51 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  52 #endif
  53 int __read_mostly nmi_watchdog_enabled;
  54 int __read_mostly soft_watchdog_enabled;
  55 int __read_mostly watchdog_user_enabled;
  56 int __read_mostly watchdog_thresh = 10;
  57
  58 #ifdef CONFIG_SMP
  59 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  60 #else
  61 #define sysctl_softlockup_all_cpu_backtrace 0
  62 #endif
  63 static struct cpumask watchdog_cpumask __read_mostly;
  64 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  65
  66 /* Helper for online, unparked cpus. */
  67 #define for_each_watchdog_cpu(cpu) \
  68         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
  69
  70 /*
  71  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  72  * are registered/started and is set to 0 when the watchdog threads are
  73  * unregistered/stopped, so it is an indicator whether the threads exist.
  74  */
  75 static int __read_mostly watchdog_running;
  76 /*
  77  * If a subsystem has a need to deactivate the watchdog temporarily, it
  78  * can use the suspend/resume interface to achieve this. The content of
  79  * the 'watchdog_suspended' variable reflects this state. Existing threads
  80  * are parked/unparked by the lockup_detector_{suspend|resume} functions
  81  * (see comment blocks pertaining to those functions for further details).
  82  *
  83  * 'watchdog_suspended' also prevents threads from being registered/started
  84  * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
  85  * of 'watchdog_running' cannot change while the watchdog is deactivated
  86  * temporarily (see related code in 'proc' handlers).
  87  */
  88 static int __read_mostly watchdog_suspended;
  89
  90 static u64 __read_mostly sample_period;
  91
  92 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  93 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  94 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  95 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  96 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  97 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  98 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
  99 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 100 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 101 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
 102 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
 103 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 104 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
 105 #endif
 106 static unsigned long soft_lockup_nmi_warn;
 107
 108 /* boot commands */
 109 /*
 110  * Should we panic when a soft-lockup or hard-lockup occurs:
 111  */
 112 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 113 static int hardlockup_panic =
 114                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
 115 /*
 116  * We may not want to enable hard lockup detection by default in all cases,
 117  * for example when running the kernel as a guest on a hypervisor. In these
 118  * cases this function can be called to disable hard lockup detection. This
 119  * function should only be executed once by the boot processor before the
 120  * kernel command line parameters are parsed, because otherwise it is not
 121  * possible to override this in hardlockup_panic_setup().
 122  */
 123 void hardlockup_detector_disable(void)
 124 {
 125         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 126 }
 127
 128 static int __init hardlockup_panic_setup(char *str)
 129 {
 130         if (!strncmp(str, "panic", 5))
 131                 hardlockup_panic = 1;
 132         else if (!strncmp(str, "nopanic", 7))
 133                 hardlockup_panic = 0;
 134         else if (!strncmp(str, "0", 1))
 135                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 136         else if (!strncmp(str, "1", 1))
 137                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 138         return 1;
 139 }
 140 __setup("nmi_watchdog=", hardlockup_panic_setup);
 141 #endif
 142
 143 unsigned int __read_mostly softlockup_panic =
 144                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 145
 146 static int __init softlockup_panic_setup(char *str)
 147 {
 148         softlockup_panic = simple_strtoul(str, NULL, 0);
 149
 150         return 1;
 151 }
 152 __setup("softlockup_panic=", softlockup_panic_setup);
 153
 154 static int __init nowatchdog_setup(char *str)
 155 {
 156         watchdog_enabled = 0;
 157         return 1;
 158 }
 159 __setup("nowatchdog", nowatchdog_setup);
 160
 161 static int __init nosoftlockup_setup(char *str)
 162 {
 163         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 164         return 1;
 165 }
 166 __setup("nosoftlockup", nosoftlockup_setup);
 167
 168 #ifdef CONFIG_SMP
 169 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 170 {
 171         sysctl_softlockup_all_cpu_backtrace =
 172                 !!simple_strtol(str, NULL, 0);
 173         return 1;
 174 }
 175 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 176 #endif
 177
 178 /*
 179  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 180  * lockups can have false positives under extreme conditions. So we generally
 181  * want a higher threshold for soft lockups than for hard lockups. So we couple
 182  * the thresholds with a factor: we make the soft threshold twice the amount of
 183  * time the hard threshold is.
 184  */
 185 static int get_softlockup_thresh(void)
 186 {
 187         return watchdog_thresh * 2;
 188 }
 189
 190 /*
 191  * Returns seconds, approximately.  We don't need nanosecond
 192  * resolution, and we don't need to waste time with a big divide when
 193  * 2^30ns == 1.074s.
 194  */
 195 static unsigned long get_timestamp(void)
 196 {
 197         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 198 }
 199
 200 static void set_sample_period(void)
 201 {
 202         /*
 203          * convert watchdog_thresh from seconds to ns
 204          * the divide by 5 is to give hrtimer several chances (two
 205          * or three with the current relation between the soft
 206          * and hard thresholds) to increment before the
 207          * hardlockup detector generates a warning
 208          */
 209         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 210 }
 211
 212 /* Commands for resetting the watchdog */
 213 static void __touch_watchdog(void)
 214 {
 215         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 216 }
 217
 218 void touch_softlockup_watchdog(void)
 219 {
 220         /*
 221          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 222          * gets zeroed here, so use the raw_ operation.
 223          */
 224         raw_cpu_write(watchdog_touch_ts, 0);
 225 }
 226 EXPORT_SYMBOL(touch_softlockup_watchdog);
 227
 228 void touch_all_softlockup_watchdogs(void)
 229 {
 230         int cpu;
 231
 232         /*
 233          * this is done lockless
 234          * do we care if a 0 races with a timestamp?
 235          * all it means is the softlock check starts one cycle later
 236          */
 237         for_each_watchdog_cpu(cpu)
 238                 per_cpu(watchdog_touch_ts, cpu) = 0;
 239 }
 240
 241 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 242 void touch_nmi_watchdog(void)
 243 {
 244         /*
 245          * Using __raw here because some code paths have
 246          * preemption enabled.  If preemption is enabled
 247          * then interrupts should be enabled too, in which
 248          * case we shouldn't have to worry about the watchdog
 249          * going off.
 250          */
 251         raw_cpu_write(watchdog_nmi_touch, true);
 252         touch_softlockup_watchdog();
 253 }
 254 EXPORT_SYMBOL(touch_nmi_watchdog);
 255
 256 #endif
 257
 258 void touch_softlockup_watchdog_sync(void)
 259 {
 260         __this_cpu_write(softlockup_touch_sync, true);
 261         __this_cpu_write(watchdog_touch_ts, 0);
 262 }
 263
 264 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 265 /* watchdog detector functions */
 266 static int is_hardlockup(void)
 267 {
 268         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 269
 270         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 271                 return 1;
 272
 273         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 274         return 0;
 275 }
 276 #endif
 277
 278 static int is_softlockup(unsigned long touch_ts)
 279 {
 280         unsigned long now = get_timestamp();
 281
 282         if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) {
 283                 /* Warn about unreasonable delays. */
 284                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 285                         return now - touch_ts;
 286         }
 287         return 0;
 288 }
 289
 290 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 291
 292 static struct perf_event_attr wd_hw_attr = {
 293         .type           = PERF_TYPE_HARDWARE,
 294         .config         = PERF_COUNT_HW_CPU_CYCLES,
 295         .size           = sizeof(struct perf_event_attr),
 296         .pinned         = 1,
 297         .disabled       = 1,
 298 };
 299
 300 /* Callback function for perf event subsystem */
 301 static void watchdog_overflow_callback(struct perf_event *event,
 302                  struct perf_sample_data *data,
 303                  struct pt_regs *regs)
 304 {
 305         /* Ensure the watchdog never gets throttled */
 306         event->hw.interrupts = 0;
 307
 308         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 309                 __this_cpu_write(watchdog_nmi_touch, false);
 310                 return;
 311         }
 312
 313         /* check for a hardlockup
 314          * This is done by making sure our timer interrupt
 315          * is incrementing.  The timer interrupt should have
 316          * fired multiple times before we overflow'd.  If it hasn't
 317          * then this is a good indication the cpu is stuck
 318          */
 319         if (is_hardlockup()) {
 320                 int this_cpu = smp_processor_id();
 321
 322                 /* only print hardlockups once */
 323                 if (__this_cpu_read(hard_watchdog_warn) == true)
 324                         return;
 325
 326                 if (hardlockup_panic)
 327                         panic("Watchdog detected hard LOCKUP on cpu %d",
 328                               this_cpu);
 329                 else
 330                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
 331                              this_cpu);
 332
 333                 __this_cpu_write(hard_watchdog_warn, true);
 334                 return;
 335         }
 336
 337         __this_cpu_write(hard_watchdog_warn, false);
 338         return;
 339 }
 340 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 341
 342 static void watchdog_interrupt_count(void)
 343 {
 344         __this_cpu_inc(hrtimer_interrupts);
 345 }
 346
 347 static int watchdog_nmi_enable(unsigned int cpu);
 348 static void watchdog_nmi_disable(unsigned int cpu);
 349
 350 /* watchdog kicker functions */
 351 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 352 {
 353         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 354         struct pt_regs *regs = get_irq_regs();
 355         int duration;
 356         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 357
 358         /* kick the hardlockup detector */
 359         watchdog_interrupt_count();
 360
 361         /* kick the softlockup detector */
 362         wake_up_process(__this_cpu_read(softlockup_watchdog));
 363
 364         /* .. and repeat */
 365         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 366
 367         if (touch_ts == 0) {
 368                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 369                         /*
 370                          * If the time stamp was touched atomically
 371                          * make sure the scheduler tick is up to date.
 372                          */
 373                         __this_cpu_write(softlockup_touch_sync, false);
 374                         sched_clock_tick();
 375                 }
 376
 377                 /* Clear the guest paused flag on watchdog reset */
 378                 kvm_check_and_clear_guest_paused();
 379                 __touch_watchdog();
 380                 return HRTIMER_RESTART;
 381         }
 382
 383         /* check for a softlockup
 384          * This is done by making sure a high priority task is
 385          * being scheduled.  The task touches the watchdog to
 386          * indicate it is getting cpu time.  If it hasn't then
 387          * this is a good indication some task is hogging the cpu
 388          */
 389         duration = is_softlockup(touch_ts);
 390         if (unlikely(duration)) {
 391                 /*
 392                  * If a virtual machine is stopped by the host it can look to
 393                  * the watchdog like a soft lockup, check to see if the host
 394                  * stopped the vm before we issue the warning
 395                  */
 396                 if (kvm_check_and_clear_guest_paused())
 397                         return HRTIMER_RESTART;
 398
 399                 /* only warn once */
 400                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 401                         /*
 402                          * When multiple processes are causing softlockups the
 403                          * softlockup detector only warns on the first one
 404                          * because the code relies on a full quiet cycle to
 405                          * re-arm.  The second process prevents the quiet cycle
 406                          * and never gets reported.  Use task pointers to detect
 407                          * this.
 408                          */
 409                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 410                             current) {
 411                                 __this_cpu_write(soft_watchdog_warn, false);
 412                                 __touch_watchdog();
 413                         }
 414                         return HRTIMER_RESTART;
 415                 }
 416
 417                 if (softlockup_all_cpu_backtrace) {
 418                         /* Prevent multiple soft-lockup reports if one cpu is already
 419                          * engaged in dumping cpu back traces
 420                          */
 421                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 422                                 /* Someone else will report us. Let's give up */
 423                                 __this_cpu_write(soft_watchdog_warn, true);
 424                                 return HRTIMER_RESTART;
 425                         }
 426                 }
 427
 428                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 429                         smp_processor_id(), duration,
 430                         current->comm, task_pid_nr(current));
 431                 __this_cpu_write(softlockup_task_ptr_saved, current);
 432                 print_modules();
 433                 print_irqtrace_events(current);
 434                 if (regs)
 435                         show_regs(regs);
 436                 else
 437                         dump_stack();
 438
 439                 if (softlockup_all_cpu_backtrace) {
 440                         /* Avoid generating two back traces for current
 441                          * given that one is already made above
 442                          */
 443                         trigger_allbutself_cpu_backtrace();
 444
 445                         clear_bit(0, &soft_lockup_nmi_warn);
 446                         /* Barrier to sync with other cpus */
 447                         smp_mb__after_atomic();
 448                 }
 449
 450                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 451                 if (softlockup_panic)
 452                         panic("softlockup: hung tasks");
 453                 __this_cpu_write(soft_watchdog_warn, true);
 454         } else
 455                 __this_cpu_write(soft_watchdog_warn, false);
 456
 457         return HRTIMER_RESTART;
 458 }
 459
 460 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 461 {
 462         struct sched_param param = { .sched_priority = prio };
 463
 464         sched_setscheduler(current, policy, &param);
 465 }
 466
 467 static void watchdog_enable(unsigned int cpu)
 468 {
 469         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 470
 471         /* kick off the timer for the hardlockup detector */
 472         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 473         hrtimer->function = watchdog_timer_fn;
 474
 475         /* Enable the perf event */
 476         watchdog_nmi_enable(cpu);
 477
 478         /* done here because hrtimer_start can only pin to smp_processor_id() */
 479         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 480                       HRTIMER_MODE_REL_PINNED);
 481
 482         /* initialize timestamp */
 483         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 484         __touch_watchdog();
 485 }
 486
 487 static void watchdog_disable(unsigned int cpu)
 488 {
 489         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 490
 491         watchdog_set_prio(SCHED_NORMAL, 0);
 492         hrtimer_cancel(hrtimer);
 493         /* disable the perf event */
 494         watchdog_nmi_disable(cpu);
 495 }
 496
 497 static void watchdog_cleanup(unsigned int cpu, bool online)
 498 {
 499         watchdog_disable(cpu);
 500 }
 501
 502 static int watchdog_should_run(unsigned int cpu)
 503 {
 504         return __this_cpu_read(hrtimer_interrupts) !=
 505                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 506 }
 507
 508 /*
 509  * The watchdog thread function - touches the timestamp.
 510  *
 511  * It only runs once every sample_period seconds (4 seconds by
 512  * default) to reset the softlockup timestamp. If this gets delayed
 513  * for more than 2*watchdog_thresh seconds then the debug-printout
 514  * triggers in watchdog_timer_fn().
 515  */
 516 static void watchdog(unsigned int cpu)
 517 {
 518         __this_cpu_write(soft_lockup_hrtimer_cnt,
 519                          __this_cpu_read(hrtimer_interrupts));
 520         __touch_watchdog();
 521
 522         /*
 523          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 524          * failure path. Check for failures that can occur asynchronously -
 525          * for example, when CPUs are on-lined - and shut down the hardware
 526          * perf event on each CPU accordingly.
 527          *
 528          * The only non-obvious place this bit can be cleared is through
 529          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 530          * pr_info here would be too noisy as it would result in a message
 531          * every few seconds if the hardlockup was disabled but the softlockup
 532          * enabled.
 533          */
 534         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 535                 watchdog_nmi_disable(cpu);
 536 }
 537
 538 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 539 /*
 540  * People like the simple clean cpu node info on boot.
 541  * Reduce the watchdog noise by only printing messages
 542  * that are different from what cpu0 displayed.
 543  */
 544 static unsigned long cpu0_err;
 545
 546 static int watchdog_nmi_enable(unsigned int cpu)
 547 {
 548         struct perf_event_attr *wd_attr;
 549         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 550
 551         /* nothing to do if the hard lockup detector is disabled */
 552         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 553                 goto out;
 554
 555         /* is it already setup and enabled? */
 556         if (event && event->state > PERF_EVENT_STATE_OFF)
 557                 goto out;
 558
 559         /* it is setup but not enabled */
 560         if (event != NULL)
 561                 goto out_enable;
 562
 563         wd_attr = &wd_hw_attr;
 564         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 565
 566         /* Try to register using hardware perf events */
 567         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 568
 569         /* save cpu0 error for future comparision */
 570         if (cpu == 0 && IS_ERR(event))
 571                 cpu0_err = PTR_ERR(event);
 572
 573         if (!IS_ERR(event)) {
 574                 /* only print for cpu0 or different than cpu0 */
 575                 if (cpu == 0 || cpu0_err)
 576                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 577                 goto out_save;
 578         }
 579
 580         /*
 581          * Disable the hard lockup detector if _any_ CPU fails to set up
 582          * set up the hardware perf event. The watchdog() function checks
 583          * the NMI_WATCHDOG_ENABLED bit periodically.
 584          *
 585          * The barriers are for syncing up watchdog_enabled across all the
 586          * cpus, as clear_bit() does not use barriers.
 587          */
 588         smp_mb__before_atomic();
 589         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 590         smp_mb__after_atomic();
 591
 592         /* skip displaying the same error again */
 593         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 594                 return PTR_ERR(event);
 595
 596         /* vary the KERN level based on the returned errno */
 597         if (PTR_ERR(event) == -EOPNOTSUPP)
 598                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 599         else if (PTR_ERR(event) == -ENOENT)
 600                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 601                          cpu);
 602         else
 603                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 604                         cpu, PTR_ERR(event));
 605
 606         pr_info("Shutting down hard lockup detector on all cpus\n");
 607
 608         return PTR_ERR(event);
 609
 610         /* success path */
 611 out_save:
 612         per_cpu(watchdog_ev, cpu) = event;
 613 out_enable:
 614         perf_event_enable(per_cpu(watchdog_ev, cpu));
 615 out:
 616         return 0;
 617 }
 618
 619 static void watchdog_nmi_disable(unsigned int cpu)
 620 {
 621         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 622
 623         if (event) {
 624                 perf_event_disable(event);
 625                 per_cpu(watchdog_ev, cpu) = NULL;
 626
 627                 /* should be in cleanup, but blocks oprofile */
 628                 perf_event_release_kernel(event);
 629         }
 630         if (cpu == 0) {
 631                 /* watchdog_nmi_enable() expects this to be zero initially. */
 632                 cpu0_err = 0;
 633         }
 634 }
 635
 636 #else
 637 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 638 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 639 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 640
 641 static struct smp_hotplug_thread watchdog_threads = {
 642         .store                  = &softlockup_watchdog,
 643         .thread_should_run      = watchdog_should_run,
 644         .thread_fn              = watchdog,
 645         .thread_comm            = "watchdog/%u",
 646         .setup                  = watchdog_enable,
 647         .cleanup                = watchdog_cleanup,
 648         .park                   = watchdog_disable,
 649         .unpark                 = watchdog_enable,
 650 };
 651
 652 /*
 653  * park all watchdog threads that are specified in 'watchdog_cpumask'
 654  */
 655 static int watchdog_park_threads(void)
 656 {
 657         int cpu, ret = 0;
 658
 659         get_online_cpus();
 660         for_each_watchdog_cpu(cpu) {
 661                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 662                 if (ret)
 663                         break;
 664         }
 665         if (ret) {
 666                 for_each_watchdog_cpu(cpu)
 667                         kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 668         }
 669         put_online_cpus();
 670
 671         return ret;
 672 }
 673
 674 /*
 675  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 676  */
 677 static void watchdog_unpark_threads(void)
 678 {
 679         int cpu;
 680
 681         get_online_cpus();
 682         for_each_watchdog_cpu(cpu)
 683                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 684         put_online_cpus();
 685 }
 686
 687 /*
 688  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 689  */
 690 int lockup_detector_suspend(void)
 691 {
 692         int ret = 0;
 693
 694         mutex_lock(&watchdog_proc_mutex);
 695         /*
 696          * Multiple suspend requests can be active in parallel (counted by
 697          * the 'watchdog_suspended' variable). If the watchdog threads are
 698          * running, the first caller takes care that they will be parked.
 699          * The state of 'watchdog_running' cannot change while a suspend
 700          * request is active (see related code in 'proc' handlers).
 701          */
 702         if (watchdog_running && !watchdog_suspended)
 703                 ret = watchdog_park_threads();
 704
 705         if (ret == 0)
 706                 watchdog_suspended++;
 707
 708         mutex_unlock(&watchdog_proc_mutex);
 709
 710         return ret;
 711 }
 712
 713 /*
 714  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 715  */
 716 void lockup_detector_resume(void)
 717 {
 718         mutex_lock(&watchdog_proc_mutex);
 719
 720         watchdog_suspended--;
 721         /*
 722          * The watchdog threads are unparked if they were previously running
 723          * and if there is no more active suspend request.
 724          */
 725         if (watchdog_running && !watchdog_suspended)
 726                 watchdog_unpark_threads();
 727
 728         mutex_unlock(&watchdog_proc_mutex);
 729 }
 730
 731 static void update_watchdog_all_cpus(void)
 732 {
 733         watchdog_park_threads();
 734         watchdog_unpark_threads();
 735 }
 736
 737 static int watchdog_enable_all_cpus(void)
 738 {
 739         int err = 0;
 740
 741         if (!watchdog_running) {
 742                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 743                                                              &watchdog_cpumask);
 744                 if (err)
 745                         pr_err("Failed to create watchdog threads, disabled\n");
 746                 else
 747                         watchdog_running = 1;
 748         } else {
 749                 /*
 750                  * Enable/disable the lockup detectors or
 751                  * change the sample period 'on the fly'.
 752                  */
 753                 update_watchdog_all_cpus();
 754         }
 755
 756         return err;
 757 }
 758
 759 /* prepare/enable/disable routines */
 760 /* sysctl functions */
 761 #ifdef CONFIG_SYSCTL
 762 static void watchdog_disable_all_cpus(void)
 763 {
 764         if (watchdog_running) {
 765                 watchdog_running = 0;
 766                 smpboot_unregister_percpu_thread(&watchdog_threads);
 767         }
 768 }
 769
 770 /*
 771  * Update the run state of the lockup detectors.
 772  */
 773 static int proc_watchdog_update(void)
 774 {
 775         int err = 0;
 776
 777         /*
 778          * Watchdog threads won't be started if they are already active.
 779          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 780          * care of this. If those threads are already active, the sample
 781          * period will be updated and the lockup detectors will be enabled
 782          * or disabled 'on the fly'.
 783          */
 784         if (watchdog_enabled && watchdog_thresh)
 785                 err = watchdog_enable_all_cpus();
 786         else
 787                 watchdog_disable_all_cpus();
 788
 789         return err;
 790
 791 }
 792
 793 /*
 794  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 795  *
 796  * caller             | table->data points to | 'which' contains the flag(s)
 797  * -------------------|-----------------------|-----------------------------
 798  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 799  *                    |                       | with SOFT_WATCHDOG_ENABLED
 800  * -------------------|-----------------------|-----------------------------
 801  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 802  * -------------------|-----------------------|-----------------------------
 803  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 804  */
 805 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 806                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 807 {
 808         int err, old, new;
 809         int *watchdog_param = (int *)table->data;
 810
 811         mutex_lock(&watchdog_proc_mutex);
 812
 813         if (watchdog_suspended) {
 814                 /* no parameter changes allowed while watchdog is suspended */
 815                 err = -EAGAIN;
 816                 goto out;
 817         }
 818
 819         /*
 820          * If the parameter is being read return the state of the corresponding
 821          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 822          * run state of the lockup detectors.
 823          */
 824         if (!write) {
 825                 *watchdog_param = (watchdog_enabled & which) != 0;
 826                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 827         } else {
 828                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 829                 if (err)
 830                         goto out;
 831
 832                 /*
 833                  * There is a race window between fetching the current value
 834                  * from 'watchdog_enabled' and storing the new value. During
 835                  * this race window, watchdog_nmi_enable() can sneak in and
 836                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 837                  * The 'cmpxchg' detects this race and the loop retries.
 838                  */
 839                 do {
 840                         old = watchdog_enabled;
 841                         /*
 842                          * If the parameter value is not zero set the
 843                          * corresponding bit(s), else clear it(them).
 844                          */
 845                         if (*watchdog_param)
 846                                 new = old | which;
 847                         else
 848                                 new = old & ~which;
 849                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 850
 851                 /*
 852                  * Update the run state of the lockup detectors.
 853                  * Restore 'watchdog_enabled' on failure.
 854                  */
 855                 err = proc_watchdog_update();
 856                 if (err)
 857                         watchdog_enabled = old;
 858         }
 859 out:
 860         mutex_unlock(&watchdog_proc_mutex);
 861         return err;
 862 }
 863
 864 /*
 865  * /proc/sys/kernel/watchdog
 866  */
 867 int proc_watchdog(struct ctl_table *table, int write,
 868                   void __user *buffer, size_t *lenp, loff_t *ppos)
 869 {
 870         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 871                                     table, write, buffer, lenp, ppos);
 872 }
 873
 874 /*
 875  * /proc/sys/kernel/nmi_watchdog
 876  */
 877 int proc_nmi_watchdog(struct ctl_table *table, int write,
 878                       void __user *buffer, size_t *lenp, loff_t *ppos)
 879 {
 880         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 881                                     table, write, buffer, lenp, ppos);
 882 }
 883
 884 /*
 885  * /proc/sys/kernel/soft_watchdog
 886  */
 887 int proc_soft_watchdog(struct ctl_table *table, int write,
 888                         void __user *buffer, size_t *lenp, loff_t *ppos)
 889 {
 890         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 891                                     table, write, buffer, lenp, ppos);
 892 }
 893
 894 /*
 895  * /proc/sys/kernel/watchdog_thresh
 896  */
 897 int proc_watchdog_thresh(struct ctl_table *table, int write,
 898                          void __user *buffer, size_t *lenp, loff_t *ppos)
 899 {
 900         int err, old;
 901
 902         mutex_lock(&watchdog_proc_mutex);
 903
 904         if (watchdog_suspended) {
 905                 /* no parameter changes allowed while watchdog is suspended */
 906                 err = -EAGAIN;
 907                 goto out;
 908         }
 909
 910         old = ACCESS_ONCE(watchdog_thresh);
 911         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 912
 913         if (err || !write)
 914                 goto out;
 915
 916         /*
 917          * Update the sample period.
 918          * Restore 'watchdog_thresh' on failure.
 919          */
 920         set_sample_period();
 921         err = proc_watchdog_update();
 922         if (err)
 923                 watchdog_thresh = old;
 924 out:
 925         mutex_unlock(&watchdog_proc_mutex);
 926         return err;
 927 }
 928
 929 /*
 930  * The cpumask is the mask of possible cpus that the watchdog can run
 931  * on, not the mask of cpus it is actually running on.  This allows the
 932  * user to specify a mask that will include cpus that have not yet
 933  * been brought online, if desired.
 934  */
 935 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 936                           void __user *buffer, size_t *lenp, loff_t *ppos)
 937 {
 938         int err;
 939
 940         mutex_lock(&watchdog_proc_mutex);
 941
 942         if (watchdog_suspended) {
 943                 /* no parameter changes allowed while watchdog is suspended */
 944                 err = -EAGAIN;
 945                 goto out;
 946         }
 947
 948         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 949         if (!err && write) {
 950                 /* Remove impossible cpus to keep sysctl output cleaner. */
 951                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
 952                             cpu_possible_mask);
 953
 954                 if (watchdog_running) {
 955                         /*
 956                          * Failure would be due to being unable to allocate
 957                          * a temporary cpumask, so we are likely not in a
 958                          * position to do much else to make things better.
 959                          */
 960                         if (smpboot_update_cpumask_percpu_thread(
 961                                     &watchdog_threads, &watchdog_cpumask) != 0)
 962                                 pr_err("cpumask update failed\n");
 963                 }
 964         }
 965 out:
 966         mutex_unlock(&watchdog_proc_mutex);
 967         return err;
 968 }
 969
 970 #endif /* CONFIG_SYSCTL */
 971
 972 void __init lockup_detector_init(void)
 973 {
 974         set_sample_period();
 975
 976 #ifdef CONFIG_NO_HZ_FULL
 977         if (tick_nohz_full_enabled()) {
 978                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 979                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
 980         } else
 981                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 982 #else
 983         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 984 #endif
 985
 986         if (watchdog_enabled)
 987                 watchdog_enable_all_cpus();
 988 }