kernel/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 static DEFINE_MUTEX(watchdog_proc_mutex);
  45
  46 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  47 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
  48 #else
  49 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  50 #endif
  51 int __read_mostly nmi_watchdog_enabled;
  52 int __read_mostly soft_watchdog_enabled;
  53 int __read_mostly watchdog_user_enabled;
  54 int __read_mostly watchdog_thresh = 10;
  55
  56 #ifdef CONFIG_SMP
  57 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  58 #else
  59 #define sysctl_softlockup_all_cpu_backtrace 0
  60 #endif
  61
  62 static int __read_mostly watchdog_running;
  63 static u64 __read_mostly sample_period;
  64
  65 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  66 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  67 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  68 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  69 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  70 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  71 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
  72 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
  73 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  74 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  75 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  76 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  77 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  78 #endif
  79 static unsigned long soft_lockup_nmi_warn;
  80
  81 /* boot commands */
  82 /*
  83  * Should we panic when a soft-lockup or hard-lockup occurs:
  84  */
  85 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  86 static int hardlockup_panic =
  87                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  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 hardlockup_detector_disable(void)
  97 {
  98         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  99 }
 100
 101 static int __init hardlockup_panic_setup(char *str)
 102 {
 103         if (!strncmp(str, "panic", 5))
 104                 hardlockup_panic = 1;
 105         else if (!strncmp(str, "nopanic", 7))
 106                 hardlockup_panic = 0;
 107         else if (!strncmp(str, "0", 1))
 108                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 109         else if (!strncmp(str, "1", 1))
 110                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 111         return 1;
 112 }
 113 __setup("nmi_watchdog=", hardlockup_panic_setup);
 114 #endif
 115
 116 unsigned int __read_mostly softlockup_panic =
 117                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 118
 119 static int __init softlockup_panic_setup(char *str)
 120 {
 121         softlockup_panic = simple_strtoul(str, NULL, 0);
 122
 123         return 1;
 124 }
 125 __setup("softlockup_panic=", softlockup_panic_setup);
 126
 127 static int __init nowatchdog_setup(char *str)
 128 {
 129         watchdog_enabled = 0;
 130         return 1;
 131 }
 132 __setup("nowatchdog", nowatchdog_setup);
 133
 134 static int __init nosoftlockup_setup(char *str)
 135 {
 136         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 137         return 1;
 138 }
 139 __setup("nosoftlockup", nosoftlockup_setup);
 140
 141 #ifdef CONFIG_SMP
 142 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 143 {
 144         sysctl_softlockup_all_cpu_backtrace =
 145                 !!simple_strtol(str, NULL, 0);
 146         return 1;
 147 }
 148 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 149 #endif
 150
 151 /*
 152  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 153  * lockups can have false positives under extreme conditions. So we generally
 154  * want a higher threshold for soft lockups than for hard lockups. So we couple
 155  * the thresholds with a factor: we make the soft threshold twice the amount of
 156  * time the hard threshold is.
 157  */
 158 static int get_softlockup_thresh(void)
 159 {
 160         return watchdog_thresh * 2;
 161 }
 162
 163 /*
 164  * Returns seconds, approximately.  We don't need nanosecond
 165  * resolution, and we don't need to waste time with a big divide when
 166  * 2^30ns == 1.074s.
 167  */
 168 static unsigned long get_timestamp(void)
 169 {
 170         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 171 }
 172
 173 static void set_sample_period(void)
 174 {
 175         /*
 176          * convert watchdog_thresh from seconds to ns
 177          * the divide by 5 is to give hrtimer several chances (two
 178          * or three with the current relation between the soft
 179          * and hard thresholds) to increment before the
 180          * hardlockup detector generates a warning
 181          */
 182         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 183 }
 184
 185 /* Commands for resetting the watchdog */
 186 static void __touch_watchdog(void)
 187 {
 188         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 189 }
 190
 191 void touch_softlockup_watchdog(void)
 192 {
 193         /*
 194          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 195          * gets zeroed here, so use the raw_ operation.
 196          */
 197         raw_cpu_write(watchdog_touch_ts, 0);
 198 }
 199 EXPORT_SYMBOL(touch_softlockup_watchdog);
 200
 201 void touch_all_softlockup_watchdogs(void)
 202 {
 203         int cpu;
 204
 205         /*
 206          * this is done lockless
 207          * do we care if a 0 races with a timestamp?
 208          * all it means is the softlock check starts one cycle later
 209          */
 210         for_each_online_cpu(cpu)
 211                 per_cpu(watchdog_touch_ts, cpu) = 0;
 212 }
 213
 214 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 215 void touch_nmi_watchdog(void)
 216 {
 217         /*
 218          * Using __raw here because some code paths have
 219          * preemption enabled.  If preemption is enabled
 220          * then interrupts should be enabled too, in which
 221          * case we shouldn't have to worry about the watchdog
 222          * going off.
 223          */
 224         raw_cpu_write(watchdog_nmi_touch, true);
 225         touch_softlockup_watchdog();
 226 }
 227 EXPORT_SYMBOL(touch_nmi_watchdog);
 228
 229 #endif
 230
 231 void touch_softlockup_watchdog_sync(void)
 232 {
 233         __this_cpu_write(softlockup_touch_sync, true);
 234         __this_cpu_write(watchdog_touch_ts, 0);
 235 }
 236
 237 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 238 /* watchdog detector functions */
 239 static int is_hardlockup(void)
 240 {
 241         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 242
 243         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 244                 return 1;
 245
 246         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 247         return 0;
 248 }
 249 #endif
 250
 251 static int is_softlockup(unsigned long touch_ts)
 252 {
 253         unsigned long now = get_timestamp();
 254
 255         if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) {
 256                 /* Warn about unreasonable delays. */
 257                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 258                         return now - touch_ts;
 259         }
 260         return 0;
 261 }
 262
 263 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 264
 265 static DEFINE_RAW_SPINLOCK(watchdog_output_lock);
 266
 267 static struct perf_event_attr wd_hw_attr = {
 268         .type           = PERF_TYPE_HARDWARE,
 269         .config         = PERF_COUNT_HW_CPU_CYCLES,
 270         .size           = sizeof(struct perf_event_attr),
 271         .pinned         = 1,
 272         .disabled       = 1,
 273 };
 274
 275 /* Callback function for perf event subsystem */
 276 static void watchdog_overflow_callback(struct perf_event *event,
 277                  struct perf_sample_data *data,
 278                  struct pt_regs *regs)
 279 {
 280         /* Ensure the watchdog never gets throttled */
 281         event->hw.interrupts = 0;
 282
 283         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 284                 __this_cpu_write(watchdog_nmi_touch, false);
 285                 return;
 286         }
 287
 288         /* check for a hardlockup
 289          * This is done by making sure our timer interrupt
 290          * is incrementing.  The timer interrupt should have
 291          * fired multiple times before we overflow'd.  If it hasn't
 292          * then this is a good indication the cpu is stuck
 293          */
 294         if (is_hardlockup()) {
 295                 int this_cpu = smp_processor_id();
 296
 297                 /* only print hardlockups once */
 298                 if (__this_cpu_read(hard_watchdog_warn) == true)
 299                         return;
 300                 /*
 301                  * If early-printk is enabled then make sure we do not
 302                  * lock up in printk() and kill console logging:
 303                  */
 304                 printk_kill();
 305
 306                 if (hardlockup_panic) {
 307                         panic("Watchdog detected hard LOCKUP on cpu %d",
 308                               this_cpu);
 309                 } else {
 310                         raw_spin_lock(&watchdog_output_lock);
 311                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
 312                              this_cpu);
 313                         raw_spin_unlock(&watchdog_output_lock);
 314                 }
 315
 316                 __this_cpu_write(hard_watchdog_warn, true);
 317                 return;
 318         }
 319
 320         __this_cpu_write(hard_watchdog_warn, false);
 321         return;
 322 }
 323 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 324
 325 static void watchdog_interrupt_count(void)
 326 {
 327         __this_cpu_inc(hrtimer_interrupts);
 328 }
 329
 330 static int watchdog_nmi_enable(unsigned int cpu);
 331 static void watchdog_nmi_disable(unsigned int cpu);
 332
 333 /* watchdog kicker functions */
 334 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 335 {
 336         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 337         struct pt_regs *regs = get_irq_regs();
 338         int duration;
 339         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 340
 341         /* kick the hardlockup detector */
 342         watchdog_interrupt_count();
 343
 344         /* kick the softlockup detector */
 345         wake_up_process(__this_cpu_read(softlockup_watchdog));
 346
 347         /* .. and repeat */
 348         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 349
 350         if (touch_ts == 0) {
 351                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 352                         /*
 353                          * If the time stamp was touched atomically
 354                          * make sure the scheduler tick is up to date.
 355                          */
 356                         __this_cpu_write(softlockup_touch_sync, false);
 357                         sched_clock_tick();
 358                 }
 359
 360                 /* Clear the guest paused flag on watchdog reset */
 361                 kvm_check_and_clear_guest_paused();
 362                 __touch_watchdog();
 363                 return HRTIMER_RESTART;
 364         }
 365
 366         /* check for a softlockup
 367          * This is done by making sure a high priority task is
 368          * being scheduled.  The task touches the watchdog to
 369          * indicate it is getting cpu time.  If it hasn't then
 370          * this is a good indication some task is hogging the cpu
 371          */
 372         duration = is_softlockup(touch_ts);
 373         if (unlikely(duration)) {
 374                 /*
 375                  * If a virtual machine is stopped by the host it can look to
 376                  * the watchdog like a soft lockup, check to see if the host
 377                  * stopped the vm before we issue the warning
 378                  */
 379                 if (kvm_check_and_clear_guest_paused())
 380                         return HRTIMER_RESTART;
 381
 382                 /* only warn once */
 383                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 384                         /*
 385                          * When multiple processes are causing softlockups the
 386                          * softlockup detector only warns on the first one
 387                          * because the code relies on a full quiet cycle to
 388                          * re-arm.  The second process prevents the quiet cycle
 389                          * and never gets reported.  Use task pointers to detect
 390                          * this.
 391                          */
 392                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 393                             current) {
 394                                 __this_cpu_write(soft_watchdog_warn, false);
 395                                 __touch_watchdog();
 396                         }
 397                         return HRTIMER_RESTART;
 398                 }
 399
 400                 if (softlockup_all_cpu_backtrace) {
 401                         /* Prevent multiple soft-lockup reports if one cpu is already
 402                          * engaged in dumping cpu back traces
 403                          */
 404                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 405                                 /* Someone else will report us. Let's give up */
 406                                 __this_cpu_write(soft_watchdog_warn, true);
 407                                 return HRTIMER_RESTART;
 408                         }
 409                 }
 410
 411                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 412                         smp_processor_id(), duration,
 413                         current->comm, task_pid_nr(current));
 414                 __this_cpu_write(softlockup_task_ptr_saved, current);
 415                 print_modules();
 416                 print_irqtrace_events(current);
 417                 if (regs)
 418                         show_regs(regs);
 419                 else
 420                         dump_stack();
 421
 422                 if (softlockup_all_cpu_backtrace) {
 423                         /* Avoid generating two back traces for current
 424                          * given that one is already made above
 425                          */
 426                         trigger_allbutself_cpu_backtrace();
 427
 428                         clear_bit(0, &soft_lockup_nmi_warn);
 429                         /* Barrier to sync with other cpus */
 430                         smp_mb__after_atomic();
 431                 }
 432
 433                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 434                 if (softlockup_panic)
 435                         panic("softlockup: hung tasks");
 436                 __this_cpu_write(soft_watchdog_warn, true);
 437         } else
 438                 __this_cpu_write(soft_watchdog_warn, false);
 439
 440         return HRTIMER_RESTART;
 441 }
 442
 443 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 444 {
 445         struct sched_param param = { .sched_priority = prio };
 446
 447         sched_setscheduler(current, policy, &param);
 448 }
 449
 450 static void watchdog_enable(unsigned int cpu)
 451 {
 452         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 453
 454         /* kick off the timer for the hardlockup detector */
 455         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 456         hrtimer->function = watchdog_timer_fn;
 457         hrtimer->irqsafe = 1;
 458
 459         /* Enable the perf event */
 460         watchdog_nmi_enable(cpu);
 461
 462         /* done here because hrtimer_start can only pin to smp_processor_id() */
 463         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 464                       HRTIMER_MODE_REL_PINNED);
 465
 466         /* initialize timestamp */
 467         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 468         __touch_watchdog();
 469 }
 470
 471 static void watchdog_disable(unsigned int cpu)
 472 {
 473         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 474
 475         watchdog_set_prio(SCHED_NORMAL, 0);
 476         hrtimer_cancel(hrtimer);
 477         /* disable the perf event */
 478         watchdog_nmi_disable(cpu);
 479 }
 480
 481 static void watchdog_cleanup(unsigned int cpu, bool online)
 482 {
 483         watchdog_disable(cpu);
 484 }
 485
 486 static int watchdog_should_run(unsigned int cpu)
 487 {
 488         return __this_cpu_read(hrtimer_interrupts) !=
 489                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 490 }
 491
 492 /*
 493  * The watchdog thread function - touches the timestamp.
 494  *
 495  * It only runs once every sample_period seconds (4 seconds by
 496  * default) to reset the softlockup timestamp. If this gets delayed
 497  * for more than 2*watchdog_thresh seconds then the debug-printout
 498  * triggers in watchdog_timer_fn().
 499  */
 500 static void watchdog(unsigned int cpu)
 501 {
 502         __this_cpu_write(soft_lockup_hrtimer_cnt,
 503                          __this_cpu_read(hrtimer_interrupts));
 504         __touch_watchdog();
 505
 506         /*
 507          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 508          * failure path. Check for failures that can occur asynchronously -
 509          * for example, when CPUs are on-lined - and shut down the hardware
 510          * perf event on each CPU accordingly.
 511          *
 512          * The only non-obvious place this bit can be cleared is through
 513          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 514          * pr_info here would be too noisy as it would result in a message
 515          * every few seconds if the hardlockup was disabled but the softlockup
 516          * enabled.
 517          */
 518         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 519                 watchdog_nmi_disable(cpu);
 520 }
 521
 522 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 523 /*
 524  * People like the simple clean cpu node info on boot.
 525  * Reduce the watchdog noise by only printing messages
 526  * that are different from what cpu0 displayed.
 527  */
 528 static unsigned long cpu0_err;
 529
 530 static int watchdog_nmi_enable(unsigned int cpu)
 531 {
 532         struct perf_event_attr *wd_attr;
 533         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 534
 535         /* nothing to do if the hard lockup detector is disabled */
 536         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 537                 goto out;
 538
 539         /* is it already setup and enabled? */
 540         if (event && event->state > PERF_EVENT_STATE_OFF)
 541                 goto out;
 542
 543         /* it is setup but not enabled */
 544         if (event != NULL)
 545                 goto out_enable;
 546
 547         wd_attr = &wd_hw_attr;
 548         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 549
 550         /* Try to register using hardware perf events */
 551         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 552
 553         /* save cpu0 error for future comparision */
 554         if (cpu == 0 && IS_ERR(event))
 555                 cpu0_err = PTR_ERR(event);
 556
 557         if (!IS_ERR(event)) {
 558                 /* only print for cpu0 or different than cpu0 */
 559                 if (cpu == 0 || cpu0_err)
 560                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 561                 goto out_save;
 562         }
 563
 564         /*
 565          * Disable the hard lockup detector if _any_ CPU fails to set up
 566          * set up the hardware perf event. The watchdog() function checks
 567          * the NMI_WATCHDOG_ENABLED bit periodically.
 568          *
 569          * The barriers are for syncing up watchdog_enabled across all the
 570          * cpus, as clear_bit() does not use barriers.
 571          */
 572         smp_mb__before_atomic();
 573         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 574         smp_mb__after_atomic();
 575
 576         /* skip displaying the same error again */
 577         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 578                 return PTR_ERR(event);
 579
 580         /* vary the KERN level based on the returned errno */
 581         if (PTR_ERR(event) == -EOPNOTSUPP)
 582                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 583         else if (PTR_ERR(event) == -ENOENT)
 584                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 585                          cpu);
 586         else
 587                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 588                         cpu, PTR_ERR(event));
 589
 590         pr_info("Shutting down hard lockup detector on all cpus\n");
 591
 592         return PTR_ERR(event);
 593
 594         /* success path */
 595 out_save:
 596         per_cpu(watchdog_ev, cpu) = event;
 597 out_enable:
 598         perf_event_enable(per_cpu(watchdog_ev, cpu));
 599 out:
 600         return 0;
 601 }
 602
 603 static void watchdog_nmi_disable(unsigned int cpu)
 604 {
 605         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 606
 607         if (event) {
 608                 perf_event_disable(event);
 609                 per_cpu(watchdog_ev, cpu) = NULL;
 610
 611                 /* should be in cleanup, but blocks oprofile */
 612                 perf_event_release_kernel(event);
 613         }
 614         if (cpu == 0) {
 615                 /* watchdog_nmi_enable() expects this to be zero initially. */
 616                 cpu0_err = 0;
 617         }
 618 }
 619
 620 void watchdog_nmi_enable_all(void)
 621 {
 622         int cpu;
 623
 624         mutex_lock(&watchdog_proc_mutex);
 625
 626         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 627                 goto unlock;
 628
 629         get_online_cpus();
 630         for_each_online_cpu(cpu)
 631                 watchdog_nmi_enable(cpu);
 632         put_online_cpus();
 633
 634 unlock:
 635         mutex_unlock(&watchdog_proc_mutex);
 636 }
 637
 638 void watchdog_nmi_disable_all(void)
 639 {
 640         int cpu;
 641
 642         mutex_lock(&watchdog_proc_mutex);
 643
 644         if (!watchdog_running)
 645                 goto unlock;
 646
 647         get_online_cpus();
 648         for_each_online_cpu(cpu)
 649                 watchdog_nmi_disable(cpu);
 650         put_online_cpus();
 651
 652 unlock:
 653         mutex_unlock(&watchdog_proc_mutex);
 654 }
 655 #else
 656 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 657 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 658 void watchdog_nmi_enable_all(void) {}
 659 void watchdog_nmi_disable_all(void) {}
 660 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 661
 662 static struct smp_hotplug_thread watchdog_threads = {
 663         .store                  = &softlockup_watchdog,
 664         .thread_should_run      = watchdog_should_run,
 665         .thread_fn              = watchdog,
 666         .thread_comm            = "watchdog/%u",
 667         .setup                  = watchdog_enable,
 668         .cleanup                = watchdog_cleanup,
 669         .park                   = watchdog_disable,
 670         .unpark                 = watchdog_enable,
 671 };
 672
 673 static void restart_watchdog_hrtimer(void *info)
 674 {
 675         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 676         int ret;
 677
 678         /*
 679          * No need to cancel and restart hrtimer if it is currently executing
 680          * because it will reprogram itself with the new period now.
 681          * We should never see it unqueued here because we are running per-cpu
 682          * with interrupts disabled.
 683          */
 684         ret = hrtimer_try_to_cancel(hrtimer);
 685         if (ret == 1)
 686                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 687                                 HRTIMER_MODE_REL_PINNED);
 688 }
 689
 690 static void update_watchdog(int cpu)
 691 {
 692         /*
 693          * Make sure that perf event counter will adopt to a new
 694          * sampling period. Updating the sampling period directly would
 695          * be much nicer but we do not have an API for that now so
 696          * let's use a big hammer.
 697          * Hrtimer will adopt the new period on the next tick but this
 698          * might be late already so we have to restart the timer as well.
 699          */
 700         watchdog_nmi_disable(cpu);
 701         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 702         watchdog_nmi_enable(cpu);
 703 }
 704
 705 static void update_watchdog_all_cpus(void)
 706 {
 707         int cpu;
 708
 709         get_online_cpus();
 710         for_each_online_cpu(cpu)
 711                 update_watchdog(cpu);
 712         put_online_cpus();
 713 }
 714
 715 static int watchdog_enable_all_cpus(void)
 716 {
 717         int err = 0;
 718
 719         if (!watchdog_running) {
 720                 err = smpboot_register_percpu_thread(&watchdog_threads);
 721                 if (err)
 722                         pr_err("Failed to create watchdog threads, disabled\n");
 723                 else
 724                         watchdog_running = 1;
 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 #endif /* CONFIG_SYSCTL */
 894
 895 void __init lockup_detector_init(void)
 896 {
 897         set_sample_period();
 898
 899         if (watchdog_enabled)
 900                 watchdog_enable_all_cpus();
 901 }