kernel/watchdog.c

   1 /*
   2  * Detect hard and soft lockups on a system
   3  *
   4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   5  *
   6  * Note: Most of this code is borrowed heavily from the original softlockup
   7  * detector, so thanks to Ingo for the initial implementation.
   8  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
   9  * to those contributors as well.
  10  */
  11
  12 #define pr_fmt(fmt) "NMI watchdog: " fmt
  13
  14 #include <linux/mm.h>
  15 #include <linux/cpu.h>
  16 #include <linux/nmi.h>
  17 #include <linux/init.h>
  18 #include <linux/module.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/smpboot.h>
  21 #include <linux/sched/rt.h>
  22
  23 #include <asm/irq_regs.h>
  24 #include <linux/kvm_para.h>
  25 #include <linux/perf_event.h>
  26
  27 /*
  28  * The run state of the lockup detectors is controlled by the content of the
  29  * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
  30  * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
  31  *
  32  * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
  33  * are variables that are only used as an 'interface' between the parameters
  34  * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
  35  * 'watchdog_thresh' variable is handled differently because its value is not
  36  * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
  37  * is equal zero.
  38  */
  39 #define NMI_WATCHDOG_ENABLED_BIT   0
  40 #define SOFT_WATCHDOG_ENABLED_BIT  1
  41 #define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
  42 #define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
  43
  44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  45 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
  46 #else
  47 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  48 #endif
  49 int __read_mostly nmi_watchdog_enabled;
  50 int __read_mostly soft_watchdog_enabled;
  51 int __read_mostly watchdog_user_enabled;
  52 int __read_mostly watchdog_thresh = 10;
  53
  54 #ifdef CONFIG_SMP
  55 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  56 #else
  57 #define sysctl_softlockup_all_cpu_backtrace 0
  58 #endif
  59
  60 static int __read_mostly watchdog_running;
  61 static u64 __read_mostly sample_period;
  62
  63 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  64 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  65 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  66 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  67 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  68 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  69 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
  70 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
  71 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  72 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  73 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  74 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  75 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  76 #endif
  77 static unsigned long soft_lockup_nmi_warn;
  78
  79 /* boot commands */
  80 /*
  81  * Should we panic when a soft-lockup or hard-lockup occurs:
  82  */
  83 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  84 static int hardlockup_panic =
  85                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  86
  87 static bool hardlockup_detector_enabled = true;
  88 /*
  89  * We may not want to enable hard lockup detection by default in all cases,
  90  * for example when running the kernel as a guest on a hypervisor. In these
  91  * cases this function can be called to disable hard lockup detection. This
  92  * function should only be executed once by the boot processor before the
  93  * kernel command line parameters are parsed, because otherwise it is not
  94  * possible to override this in hardlockup_panic_setup().
  95  */
  96 void watchdog_enable_hardlockup_detector(bool val)
  97 {
  98         hardlockup_detector_enabled = val;
  99 }
 100
 101 bool watchdog_hardlockup_detector_is_enabled(void)
 102 {
 103         return hardlockup_detector_enabled;
 104 }
 105
 106 static int __init hardlockup_panic_setup(char *str)
 107 {
 108         if (!strncmp(str, "panic", 5))
 109                 hardlockup_panic = 1;
 110         else if (!strncmp(str, "nopanic", 7))
 111                 hardlockup_panic = 0;
 112         else if (!strncmp(str, "0", 1))
 113                 watchdog_user_enabled = 0;
 114         else if (!strncmp(str, "1", 1) || !strncmp(str, "2", 1)) {
 115                 /*
 116                  * Setting 'nmi_watchdog=1' or 'nmi_watchdog=2' (legacy option)
 117                  * has the same effect.
 118                  */
 119                 watchdog_user_enabled = 1;
 120                 watchdog_enable_hardlockup_detector(true);
 121         }
 122         return 1;
 123 }
 124 __setup("nmi_watchdog=", hardlockup_panic_setup);
 125 #endif
 126
 127 unsigned int __read_mostly softlockup_panic =
 128                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 129
 130 static int __init softlockup_panic_setup(char *str)
 131 {
 132         softlockup_panic = simple_strtoul(str, NULL, 0);
 133
 134         return 1;
 135 }
 136 __setup("softlockup_panic=", softlockup_panic_setup);
 137
 138 static int __init nowatchdog_setup(char *str)
 139 {
 140         watchdog_user_enabled = 0;
 141         return 1;
 142 }
 143 __setup("nowatchdog", nowatchdog_setup);
 144
 145 /* deprecated */
 146 static int __init nosoftlockup_setup(char *str)
 147 {
 148         watchdog_user_enabled = 0;
 149         return 1;
 150 }
 151 __setup("nosoftlockup", nosoftlockup_setup);
 152 /*  */
 153 #ifdef CONFIG_SMP
 154 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 155 {
 156         sysctl_softlockup_all_cpu_backtrace =
 157                 !!simple_strtol(str, NULL, 0);
 158         return 1;
 159 }
 160 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 161 #endif
 162
 163 /*
 164  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 165  * lockups can have false positives under extreme conditions. So we generally
 166  * want a higher threshold for soft lockups than for hard lockups. So we couple
 167  * the thresholds with a factor: we make the soft threshold twice the amount of
 168  * time the hard threshold is.
 169  */
 170 static int get_softlockup_thresh(void)
 171 {
 172         return watchdog_thresh * 2;
 173 }
 174
 175 /*
 176  * Returns seconds, approximately.  We don't need nanosecond
 177  * resolution, and we don't need to waste time with a big divide when
 178  * 2^30ns == 1.074s.
 179  */
 180 static unsigned long get_timestamp(void)
 181 {
 182         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 183 }
 184
 185 static void set_sample_period(void)
 186 {
 187         /*
 188          * convert watchdog_thresh from seconds to ns
 189          * the divide by 5 is to give hrtimer several chances (two
 190          * or three with the current relation between the soft
 191          * and hard thresholds) to increment before the
 192          * hardlockup detector generates a warning
 193          */
 194         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 195 }
 196
 197 /* Commands for resetting the watchdog */
 198 static void __touch_watchdog(void)
 199 {
 200         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 201 }
 202
 203 void touch_softlockup_watchdog(void)
 204 {
 205         /*
 206          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 207          * gets zeroed here, so use the raw_ operation.
 208          */
 209         raw_cpu_write(watchdog_touch_ts, 0);
 210 }
 211 EXPORT_SYMBOL(touch_softlockup_watchdog);
 212
 213 void touch_all_softlockup_watchdogs(void)
 214 {
 215         int cpu;
 216
 217         /*
 218          * this is done lockless
 219          * do we care if a 0 races with a timestamp?
 220          * all it means is the softlock check starts one cycle later
 221          */
 222         for_each_online_cpu(cpu)
 223                 per_cpu(watchdog_touch_ts, cpu) = 0;
 224 }
 225
 226 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 227 void touch_nmi_watchdog(void)
 228 {
 229         /*
 230          * Using __raw here because some code paths have
 231          * preemption enabled.  If preemption is enabled
 232          * then interrupts should be enabled too, in which
 233          * case we shouldn't have to worry about the watchdog
 234          * going off.
 235          */
 236         raw_cpu_write(watchdog_nmi_touch, true);
 237         touch_softlockup_watchdog();
 238 }
 239 EXPORT_SYMBOL(touch_nmi_watchdog);
 240
 241 #endif
 242
 243 void touch_softlockup_watchdog_sync(void)
 244 {
 245         __this_cpu_write(softlockup_touch_sync, true);
 246         __this_cpu_write(watchdog_touch_ts, 0);
 247 }
 248
 249 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 250 /* watchdog detector functions */
 251 static int is_hardlockup(void)
 252 {
 253         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 254
 255         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 256                 return 1;
 257
 258         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 259         return 0;
 260 }
 261 #endif
 262
 263 static int is_softlockup(unsigned long touch_ts)
 264 {
 265         unsigned long now = get_timestamp();
 266
 267         /* Warn about unreasonable delays: */
 268         if (time_after(now, touch_ts + get_softlockup_thresh()))
 269                 return now - touch_ts;
 270
 271         return 0;
 272 }
 273
 274 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 275
 276 static struct perf_event_attr wd_hw_attr = {
 277         .type           = PERF_TYPE_HARDWARE,
 278         .config         = PERF_COUNT_HW_CPU_CYCLES,
 279         .size           = sizeof(struct perf_event_attr),
 280         .pinned         = 1,
 281         .disabled       = 1,
 282 };
 283
 284 /* Callback function for perf event subsystem */
 285 static void watchdog_overflow_callback(struct perf_event *event,
 286                  struct perf_sample_data *data,
 287                  struct pt_regs *regs)
 288 {
 289         /* Ensure the watchdog never gets throttled */
 290         event->hw.interrupts = 0;
 291
 292         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 293                 __this_cpu_write(watchdog_nmi_touch, false);
 294                 return;
 295         }
 296
 297         /* check for a hardlockup
 298          * This is done by making sure our timer interrupt
 299          * is incrementing.  The timer interrupt should have
 300          * fired multiple times before we overflow'd.  If it hasn't
 301          * then this is a good indication the cpu is stuck
 302          */
 303         if (is_hardlockup()) {
 304                 int this_cpu = smp_processor_id();
 305
 306                 /* only print hardlockups once */
 307                 if (__this_cpu_read(hard_watchdog_warn) == true)
 308                         return;
 309
 310                 if (hardlockup_panic)
 311                         panic("Watchdog detected hard LOCKUP on cpu %d",
 312                               this_cpu);
 313                 else
 314                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
 315                              this_cpu);
 316
 317                 __this_cpu_write(hard_watchdog_warn, true);
 318                 return;
 319         }
 320
 321         __this_cpu_write(hard_watchdog_warn, false);
 322         return;
 323 }
 324 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 325
 326 static void watchdog_interrupt_count(void)
 327 {
 328         __this_cpu_inc(hrtimer_interrupts);
 329 }
 330
 331 static int watchdog_nmi_enable(unsigned int cpu);
 332 static void watchdog_nmi_disable(unsigned int cpu);
 333
 334 /* watchdog kicker functions */
 335 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 336 {
 337         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 338         struct pt_regs *regs = get_irq_regs();
 339         int duration;
 340         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 341
 342         /* kick the hardlockup detector */
 343         watchdog_interrupt_count();
 344
 345         /* kick the softlockup detector */
 346         wake_up_process(__this_cpu_read(softlockup_watchdog));
 347
 348         /* .. and repeat */
 349         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 350
 351         if (touch_ts == 0) {
 352                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 353                         /*
 354                          * If the time stamp was touched atomically
 355                          * make sure the scheduler tick is up to date.
 356                          */
 357                         __this_cpu_write(softlockup_touch_sync, false);
 358                         sched_clock_tick();
 359                 }
 360
 361                 /* Clear the guest paused flag on watchdog reset */
 362                 kvm_check_and_clear_guest_paused();
 363                 __touch_watchdog();
 364                 return HRTIMER_RESTART;
 365         }
 366
 367         /* check for a softlockup
 368          * This is done by making sure a high priority task is
 369          * being scheduled.  The task touches the watchdog to
 370          * indicate it is getting cpu time.  If it hasn't then
 371          * this is a good indication some task is hogging the cpu
 372          */
 373         duration = is_softlockup(touch_ts);
 374         if (unlikely(duration)) {
 375                 /*
 376                  * If a virtual machine is stopped by the host it can look to
 377                  * the watchdog like a soft lockup, check to see if the host
 378                  * stopped the vm before we issue the warning
 379                  */
 380                 if (kvm_check_and_clear_guest_paused())
 381                         return HRTIMER_RESTART;
 382
 383                 /* only warn once */
 384                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 385                         /*
 386                          * When multiple processes are causing softlockups the
 387                          * softlockup detector only warns on the first one
 388                          * because the code relies on a full quiet cycle to
 389                          * re-arm.  The second process prevents the quiet cycle
 390                          * and never gets reported.  Use task pointers to detect
 391                          * this.
 392                          */
 393                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 394                             current) {
 395                                 __this_cpu_write(soft_watchdog_warn, false);
 396                                 __touch_watchdog();
 397                         }
 398                         return HRTIMER_RESTART;
 399                 }
 400
 401                 if (softlockup_all_cpu_backtrace) {
 402                         /* Prevent multiple soft-lockup reports if one cpu is already
 403                          * engaged in dumping cpu back traces
 404                          */
 405                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 406                                 /* Someone else will report us. Let's give up */
 407                                 __this_cpu_write(soft_watchdog_warn, true);
 408                                 return HRTIMER_RESTART;
 409                         }
 410                 }
 411
 412                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 413                         smp_processor_id(), duration,
 414                         current->comm, task_pid_nr(current));
 415                 __this_cpu_write(softlockup_task_ptr_saved, current);
 416                 print_modules();
 417                 print_irqtrace_events(current);
 418                 if (regs)
 419                         show_regs(regs);
 420                 else
 421                         dump_stack();
 422
 423                 if (softlockup_all_cpu_backtrace) {
 424                         /* Avoid generating two back traces for current
 425                          * given that one is already made above
 426                          */
 427                         trigger_allbutself_cpu_backtrace();
 428
 429                         clear_bit(0, &soft_lockup_nmi_warn);
 430                         /* Barrier to sync with other cpus */
 431                         smp_mb__after_atomic();
 432                 }
 433
 434                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 435                 if (softlockup_panic)
 436                         panic("softlockup: hung tasks");
 437                 __this_cpu_write(soft_watchdog_warn, true);
 438         } else
 439                 __this_cpu_write(soft_watchdog_warn, false);
 440
 441         return HRTIMER_RESTART;
 442 }
 443
 444 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 445 {
 446         struct sched_param param = { .sched_priority = prio };
 447
 448         sched_setscheduler(current, policy, &param);
 449 }
 450
 451 static void watchdog_enable(unsigned int cpu)
 452 {
 453         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 454
 455         /* kick off the timer for the hardlockup detector */
 456         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 457         hrtimer->function = watchdog_timer_fn;
 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         /*
 536          * Some kernels need to default hard lockup detection to
 537          * 'disabled', for example a guest on a hypervisor.
 538          */
 539         if (!watchdog_hardlockup_detector_is_enabled()) {
 540                 event = ERR_PTR(-ENOENT);
 541                 goto handle_err;
 542         }
 543
 544         /* is it already setup and enabled? */
 545         if (event && event->state > PERF_EVENT_STATE_OFF)
 546                 goto out;
 547
 548         /* it is setup but not enabled */
 549         if (event != NULL)
 550                 goto out_enable;
 551
 552         wd_attr = &wd_hw_attr;
 553         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 554
 555         /* Try to register using hardware perf events */
 556         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 557
 558 handle_err:
 559         /* save cpu0 error for future comparision */
 560         if (cpu == 0 && IS_ERR(event))
 561                 cpu0_err = PTR_ERR(event);
 562
 563         if (!IS_ERR(event)) {
 564                 /* only print for cpu0 or different than cpu0 */
 565                 if (cpu == 0 || cpu0_err)
 566                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 567                 goto out_save;
 568         }
 569
 570         /*
 571          * Disable the hard lockup detector if _any_ CPU fails to set up
 572          * set up the hardware perf event. The watchdog() function checks
 573          * the NMI_WATCHDOG_ENABLED bit periodically.
 574          *
 575          * The barriers are for syncing up watchdog_enabled across all the
 576          * cpus, as clear_bit() does not use barriers.
 577          */
 578         smp_mb__before_atomic();
 579         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 580         smp_mb__after_atomic();
 581
 582         /* skip displaying the same error again */
 583         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 584                 return PTR_ERR(event);
 585
 586         /* vary the KERN level based on the returned errno */
 587         if (PTR_ERR(event) == -EOPNOTSUPP)
 588                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 589         else if (PTR_ERR(event) == -ENOENT)
 590                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 591                          cpu);
 592         else
 593                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 594                         cpu, PTR_ERR(event));
 595
 596         pr_info("Shutting down hard lockup detector on all cpus\n");
 597
 598         return PTR_ERR(event);
 599
 600         /* success path */
 601 out_save:
 602         per_cpu(watchdog_ev, cpu) = event;
 603 out_enable:
 604         perf_event_enable(per_cpu(watchdog_ev, cpu));
 605 out:
 606         return 0;
 607 }
 608
 609 static void watchdog_nmi_disable(unsigned int cpu)
 610 {
 611         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 612
 613         if (event) {
 614                 perf_event_disable(event);
 615                 per_cpu(watchdog_ev, cpu) = NULL;
 616
 617                 /* should be in cleanup, but blocks oprofile */
 618                 perf_event_release_kernel(event);
 619         }
 620         if (cpu == 0) {
 621                 /* watchdog_nmi_enable() expects this to be zero initially. */
 622                 cpu0_err = 0;
 623         }
 624 }
 625 #else
 626 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 627 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 628 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 629
 630 static struct smp_hotplug_thread watchdog_threads = {
 631         .store                  = &softlockup_watchdog,
 632         .thread_should_run      = watchdog_should_run,
 633         .thread_fn              = watchdog,
 634         .thread_comm            = "watchdog/%u",
 635         .setup                  = watchdog_enable,
 636         .cleanup                = watchdog_cleanup,
 637         .park                   = watchdog_disable,
 638         .unpark                 = watchdog_enable,
 639 };
 640
 641 static void restart_watchdog_hrtimer(void *info)
 642 {
 643         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 644         int ret;
 645
 646         /*
 647          * No need to cancel and restart hrtimer if it is currently executing
 648          * because it will reprogram itself with the new period now.
 649          * We should never see it unqueued here because we are running per-cpu
 650          * with interrupts disabled.
 651          */
 652         ret = hrtimer_try_to_cancel(hrtimer);
 653         if (ret == 1)
 654                 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 655                                 HRTIMER_MODE_REL_PINNED);
 656 }
 657
 658 static void update_timers(int cpu)
 659 {
 660         /*
 661          * Make sure that perf event counter will adopt to a new
 662          * sampling period. Updating the sampling period directly would
 663          * be much nicer but we do not have an API for that now so
 664          * let's use a big hammer.
 665          * Hrtimer will adopt the new period on the next tick but this
 666          * might be late already so we have to restart the timer as well.
 667          */
 668         watchdog_nmi_disable(cpu);
 669         smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
 670         watchdog_nmi_enable(cpu);
 671 }
 672
 673 static void update_timers_all_cpus(void)
 674 {
 675         int cpu;
 676
 677         get_online_cpus();
 678         for_each_online_cpu(cpu)
 679                 update_timers(cpu);
 680         put_online_cpus();
 681 }
 682
 683 static int watchdog_enable_all_cpus(bool sample_period_changed)
 684 {
 685         int err = 0;
 686
 687         if (!watchdog_running) {
 688                 err = smpboot_register_percpu_thread(&watchdog_threads);
 689                 if (err)
 690                         pr_err("Failed to create watchdog threads, disabled\n");
 691                 else
 692                         watchdog_running = 1;
 693         } else if (sample_period_changed) {
 694                 update_timers_all_cpus();
 695         }
 696
 697         return err;
 698 }
 699
 700 /* prepare/enable/disable routines */
 701 /* sysctl functions */
 702 #ifdef CONFIG_SYSCTL
 703 static void watchdog_disable_all_cpus(void)
 704 {
 705         if (watchdog_running) {
 706                 watchdog_running = 0;
 707                 smpboot_unregister_percpu_thread(&watchdog_threads);
 708         }
 709 }
 710
 711 /*
 712  * Update the run state of the lockup detectors.
 713  */
 714 static int proc_watchdog_update(void)
 715 {
 716         int err = 0;
 717
 718         /*
 719          * Watchdog threads won't be started if they are already active.
 720          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 721          * care of this. If those threads are already active, the sample
 722          * period will be updated and the lockup detectors will be enabled
 723          * or disabled 'on the fly'.
 724          */
 725         if (watchdog_enabled && watchdog_thresh)
 726                 err = watchdog_enable_all_cpus(true);
 727         else
 728                 watchdog_disable_all_cpus();
 729
 730         return err;
 731
 732 }
 733
 734 static DEFINE_MUTEX(watchdog_proc_mutex);
 735
 736 /*
 737  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 738  *
 739  * caller             | table->data points to | 'which' contains the flag(s)
 740  * -------------------|-----------------------|-----------------------------
 741  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 742  *                    |                       | with SOFT_WATCHDOG_ENABLED
 743  * -------------------|-----------------------|-----------------------------
 744  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 745  * -------------------|-----------------------|-----------------------------
 746  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 747  */
 748 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 749                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 750 {
 751         int err, old, new;
 752         int *watchdog_param = (int *)table->data;
 753
 754         mutex_lock(&watchdog_proc_mutex);
 755
 756         /*
 757          * If the parameter is being read return the state of the corresponding
 758          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 759          * run state of the lockup detectors.
 760          */
 761         if (!write) {
 762                 *watchdog_param = (watchdog_enabled & which) != 0;
 763                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 764         } else {
 765                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 766                 if (err)
 767                         goto out;
 768
 769                 /*
 770                  * There is a race window between fetching the current value
 771                  * from 'watchdog_enabled' and storing the new value. During
 772                  * this race window, watchdog_nmi_enable() can sneak in and
 773                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 774                  * The 'cmpxchg' detects this race and the loop retries.
 775                  */
 776                 do {
 777                         old = watchdog_enabled;
 778                         /*
 779                          * If the parameter value is not zero set the
 780                          * corresponding bit(s), else clear it(them).
 781                          */
 782                         if (*watchdog_param)
 783                                 new = old | which;
 784                         else
 785                                 new = old & ~which;
 786                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 787
 788                 /*
 789                  * Update the run state of the lockup detectors.
 790                  * Restore 'watchdog_enabled' on failure.
 791                  */
 792                 err = proc_watchdog_update();
 793                 if (err)
 794                         watchdog_enabled = old;
 795         }
 796 out:
 797         mutex_unlock(&watchdog_proc_mutex);
 798         return err;
 799 }
 800
 801 /*
 802  * /proc/sys/kernel/watchdog
 803  */
 804 int proc_watchdog(struct ctl_table *table, int write,
 805                   void __user *buffer, size_t *lenp, loff_t *ppos)
 806 {
 807         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 808                                     table, write, buffer, lenp, ppos);
 809 }
 810
 811 /*
 812  * /proc/sys/kernel/nmi_watchdog
 813  */
 814 int proc_nmi_watchdog(struct ctl_table *table, int write,
 815                       void __user *buffer, size_t *lenp, loff_t *ppos)
 816 {
 817         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 818                                     table, write, buffer, lenp, ppos);
 819 }
 820
 821 /*
 822  * /proc/sys/kernel/soft_watchdog
 823  */
 824 int proc_soft_watchdog(struct ctl_table *table, int write,
 825                         void __user *buffer, size_t *lenp, loff_t *ppos)
 826 {
 827         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 828                                     table, write, buffer, lenp, ppos);
 829 }
 830
 831 /*
 832  * /proc/sys/kernel/watchdog_thresh
 833  */
 834 int proc_watchdog_thresh(struct ctl_table *table, int write,
 835                          void __user *buffer, size_t *lenp, loff_t *ppos)
 836 {
 837         int err, old;
 838
 839         mutex_lock(&watchdog_proc_mutex);
 840
 841         old = ACCESS_ONCE(watchdog_thresh);
 842         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 843
 844         if (err || !write)
 845                 goto out;
 846
 847         /*
 848          * Update the sample period.
 849          * Restore 'watchdog_thresh' on failure.
 850          */
 851         set_sample_period();
 852         err = proc_watchdog_update();
 853         if (err)
 854                 watchdog_thresh = old;
 855 out:
 856         mutex_unlock(&watchdog_proc_mutex);
 857         return err;
 858 }
 859
 860 /*
 861  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 862  */
 863
 864 int proc_dowatchdog(struct ctl_table *table, int write,
 865                     void __user *buffer, size_t *lenp, loff_t *ppos)
 866 {
 867         int err, old_thresh, old_enabled;
 868         bool old_hardlockup;
 869
 870         mutex_lock(&watchdog_proc_mutex);
 871         old_thresh = ACCESS_ONCE(watchdog_thresh);
 872         old_enabled = ACCESS_ONCE(watchdog_user_enabled);
 873         old_hardlockup = watchdog_hardlockup_detector_is_enabled();
 874
 875         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 876         if (err || !write)
 877                 goto out;
 878
 879         set_sample_period();
 880         /*
 881          * Watchdog threads shouldn't be enabled if they are
 882          * disabled. The 'watchdog_running' variable check in
 883          * watchdog_*_all_cpus() function takes care of this.
 884          */
 885         if (watchdog_user_enabled && watchdog_thresh) {
 886                 /*
 887                  * Prevent a change in watchdog_thresh accidentally overriding
 888                  * the enablement of the hardlockup detector.
 889                  */
 890                 if (watchdog_user_enabled != old_enabled)
 891                         watchdog_enable_hardlockup_detector(true);
 892                 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
 893         } else
 894                 watchdog_disable_all_cpus();
 895
 896         /* Restore old values on failure */
 897         if (err) {
 898                 watchdog_thresh = old_thresh;
 899                 watchdog_user_enabled = old_enabled;
 900                 watchdog_enable_hardlockup_detector(old_hardlockup);
 901         }
 902 out:
 903         mutex_unlock(&watchdog_proc_mutex);
 904         return err;
 905 }
 906 #endif /* CONFIG_SYSCTL */
 907
 908 void __init lockup_detector_init(void)
 909 {
 910         set_sample_period();
 911
 912         if (watchdog_user_enabled)
 913                 watchdog_enable_all_cpus(false);
 914 }