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