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