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