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 #include <linux/tick.h>
  23
  24 #include <asm/irq_regs.h>
  25 #include <linux/kvm_para.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/kthread.h>
  28
  29 /*
  30  * The run state of the lockup detectors is controlled by the content of the
  31  * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
  32  * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
  33  *
  34  * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
  35  * are variables that are only used as an 'interface' between the parameters
  36  * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
  37  * 'watchdog_thresh' variable is handled differently because its value is not
  38  * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
  39  * is equal zero.
  40  */
  41 #define NMI_WATCHDOG_ENABLED_BIT   0
  42 #define SOFT_WATCHDOG_ENABLED_BIT  1
  43 #define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
  44 #define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
  45
  46 static DEFINE_MUTEX(watchdog_proc_mutex);
  47
  48 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  49 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
  50 #else
  51 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  52 #endif
  53 int __read_mostly nmi_watchdog_enabled;
  54 int __read_mostly soft_watchdog_enabled;
  55 int __read_mostly watchdog_user_enabled;
  56 int __read_mostly watchdog_thresh = 10;
  57
  58 #ifdef CONFIG_SMP
  59 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  60 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  61 #else
  62 #define sysctl_softlockup_all_cpu_backtrace 0
  63 #define sysctl_hardlockup_all_cpu_backtrace 0
  64 #endif
  65 static struct cpumask watchdog_cpumask __read_mostly;
  66 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  67
  68 /* Helper for online, unparked cpus. */
  69 #define for_each_watchdog_cpu(cpu) \
  70         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
  71
  72 /*
  73  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  74  * are registered/started and is set to 0 when the watchdog threads are
  75  * unregistered/stopped, so it is an indicator whether the threads exist.
  76  */
  77 static int __read_mostly watchdog_running;
  78 /*
  79  * If a subsystem has a need to deactivate the watchdog temporarily, it
  80  * can use the suspend/resume interface to achieve this. The content of
  81  * the 'watchdog_suspended' variable reflects this state. Existing threads
  82  * are parked/unparked by the lockup_detector_{suspend|resume} functions
  83  * (see comment blocks pertaining to those functions for further details).
  84  *
  85  * 'watchdog_suspended' also prevents threads from being registered/started
  86  * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
  87  * of 'watchdog_running' cannot change while the watchdog is deactivated
  88  * temporarily (see related code in 'proc' handlers).
  89  */
  90 static int __read_mostly watchdog_suspended;
  91
  92 static u64 __read_mostly sample_period;
  93
  94 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  95 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  96 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  97 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  98 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  99 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 100 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 101 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 102 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 103 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
 104 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
 105 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 106 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
 107 #endif
 108 static unsigned long soft_lockup_nmi_warn;
 109
 110 /* boot commands */
 111 /*
 112  * Should we panic when a soft-lockup or hard-lockup occurs:
 113  */
 114 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 115 unsigned int __read_mostly hardlockup_panic =
 116                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
 117 static unsigned long hardlockup_allcpu_dumped;
 118 /*
 119  * We may not want to enable hard lockup detection by default in all cases,
 120  * for example when running the kernel as a guest on a hypervisor. In these
 121  * cases this function can be called to disable hard lockup detection. This
 122  * function should only be executed once by the boot processor before the
 123  * kernel command line parameters are parsed, because otherwise it is not
 124  * possible to override this in hardlockup_panic_setup().
 125  */
 126 void hardlockup_detector_disable(void)
 127 {
 128         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 129 }
 130
 131 static int __init hardlockup_panic_setup(char *str)
 132 {
 133         if (!strncmp(str, "panic", 5))
 134                 hardlockup_panic = 1;
 135         else if (!strncmp(str, "nopanic", 7))
 136                 hardlockup_panic = 0;
 137         else if (!strncmp(str, "0", 1))
 138                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 139         else if (!strncmp(str, "1", 1))
 140                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 141         return 1;
 142 }
 143 __setup("nmi_watchdog=", hardlockup_panic_setup);
 144 #endif
 145
 146 unsigned int __read_mostly softlockup_panic =
 147                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 148
 149 static int __init softlockup_panic_setup(char *str)
 150 {
 151         softlockup_panic = simple_strtoul(str, NULL, 0);
 152
 153         return 1;
 154 }
 155 __setup("softlockup_panic=", softlockup_panic_setup);
 156
 157 static int __init nowatchdog_setup(char *str)
 158 {
 159         watchdog_enabled = 0;
 160         return 1;
 161 }
 162 __setup("nowatchdog", nowatchdog_setup);
 163
 164 static int __init nosoftlockup_setup(char *str)
 165 {
 166         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 167         return 1;
 168 }
 169 __setup("nosoftlockup", nosoftlockup_setup);
 170
 171 #ifdef CONFIG_SMP
 172 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 173 {
 174         sysctl_softlockup_all_cpu_backtrace =
 175                 !!simple_strtol(str, NULL, 0);
 176         return 1;
 177 }
 178 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 179 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
 180 {
 181         sysctl_hardlockup_all_cpu_backtrace =
 182                 !!simple_strtol(str, NULL, 0);
 183         return 1;
 184 }
 185 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
 186 #endif
 187
 188 /*
 189  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 190  * lockups can have false positives under extreme conditions. So we generally
 191  * want a higher threshold for soft lockups than for hard lockups. So we couple
 192  * the thresholds with a factor: we make the soft threshold twice the amount of
 193  * time the hard threshold is.
 194  */
 195 static int get_softlockup_thresh(void)
 196 {
 197         return watchdog_thresh * 2;
 198 }
 199
 200 /*
 201  * Returns seconds, approximately.  We don't need nanosecond
 202  * resolution, and we don't need to waste time with a big divide when
 203  * 2^30ns == 1.074s.
 204  */
 205 static unsigned long get_timestamp(void)
 206 {
 207         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 208 }
 209
 210 static void set_sample_period(void)
 211 {
 212         /*
 213          * convert watchdog_thresh from seconds to ns
 214          * the divide by 5 is to give hrtimer several chances (two
 215          * or three with the current relation between the soft
 216          * and hard thresholds) to increment before the
 217          * hardlockup detector generates a warning
 218          */
 219         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 220 }
 221
 222 /* Commands for resetting the watchdog */
 223 static void __touch_watchdog(void)
 224 {
 225         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 226 }
 227
 228 void touch_softlockup_watchdog(void)
 229 {
 230         /*
 231          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 232          * gets zeroed here, so use the raw_ operation.
 233          */
 234         raw_cpu_write(watchdog_touch_ts, 0);
 235 }
 236 EXPORT_SYMBOL(touch_softlockup_watchdog);
 237
 238 void touch_all_softlockup_watchdogs(void)
 239 {
 240         int cpu;
 241
 242         /*
 243          * this is done lockless
 244          * do we care if a 0 races with a timestamp?
 245          * all it means is the softlock check starts one cycle later
 246          */
 247         for_each_watchdog_cpu(cpu)
 248                 per_cpu(watchdog_touch_ts, cpu) = 0;
 249 }
 250
 251 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 252 void touch_nmi_watchdog(void)
 253 {
 254         /*
 255          * Using __raw here because some code paths have
 256          * preemption enabled.  If preemption is enabled
 257          * then interrupts should be enabled too, in which
 258          * case we shouldn't have to worry about the watchdog
 259          * going off.
 260          */
 261         raw_cpu_write(watchdog_nmi_touch, true);
 262         touch_softlockup_watchdog();
 263 }
 264 EXPORT_SYMBOL(touch_nmi_watchdog);
 265
 266 #endif
 267
 268 void touch_softlockup_watchdog_sync(void)
 269 {
 270         __this_cpu_write(softlockup_touch_sync, true);
 271         __this_cpu_write(watchdog_touch_ts, 0);
 272 }
 273
 274 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 275 /* watchdog detector functions */
 276 static bool is_hardlockup(void)
 277 {
 278         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 279
 280         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 281                 return true;
 282
 283         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 284         return false;
 285 }
 286 #endif
 287
 288 static int is_softlockup(unsigned long touch_ts)
 289 {
 290         unsigned long now = get_timestamp();
 291
 292         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 293                 /* Warn about unreasonable delays. */
 294                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 295                         return now - touch_ts;
 296         }
 297         return 0;
 298 }
 299
 300 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 301
 302 static DEFINE_RAW_SPINLOCK(watchdog_output_lock);
 303
 304 static struct perf_event_attr wd_hw_attr = {
 305         .type           = PERF_TYPE_HARDWARE,
 306         .config         = PERF_COUNT_HW_CPU_CYCLES,
 307         .size           = sizeof(struct perf_event_attr),
 308         .pinned         = 1,
 309         .disabled       = 1,
 310 };
 311
 312 /* Callback function for perf event subsystem */
 313 static void watchdog_overflow_callback(struct perf_event *event,
 314                  struct perf_sample_data *data,
 315                  struct pt_regs *regs)
 316 {
 317         /* Ensure the watchdog never gets throttled */
 318         event->hw.interrupts = 0;
 319
 320         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 321                 __this_cpu_write(watchdog_nmi_touch, false);
 322                 return;
 323         }
 324
 325         /* check for a hardlockup
 326          * This is done by making sure our timer interrupt
 327          * is incrementing.  The timer interrupt should have
 328          * fired multiple times before we overflow'd.  If it hasn't
 329          * then this is a good indication the cpu is stuck
 330          */
 331         if (is_hardlockup()) {
 332                 int this_cpu = smp_processor_id();
 333
 334                 /* only print hardlockups once */
 335                 if (__this_cpu_read(hard_watchdog_warn) == true)
 336                         return;
 337                 /*
 338                  * If early-printk is enabled then make sure we do not
 339                  * lock up in printk() and kill console logging:
 340                  */
 341                 printk_kill();
 342
 343                 raw_spin_lock(&watchdog_output_lock);
 344
 345                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 346                 print_modules();
 347                 print_irqtrace_events(current);
 348                 if (regs)
 349                         show_regs(regs);
 350                 else
 351                         dump_stack();
 352
 353                 /*
 354                  * Perform all-CPU dump only once to avoid multiple hardlockups
 355                  * generating interleaving traces
 356                  */
 357                 if (sysctl_hardlockup_all_cpu_backtrace &&
 358                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 359                         trigger_allbutself_cpu_backtrace();
 360
 361                 raw_spin_unlock(&watchdog_output_lock);
 362                 if (hardlockup_panic)
 363                         nmi_panic(regs, "Hard LOCKUP");
 364
 365                 __this_cpu_write(hard_watchdog_warn, true);
 366                 return;
 367         }
 368
 369         __this_cpu_write(hard_watchdog_warn, false);
 370         return;
 371 }
 372 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 373
 374 static void watchdog_interrupt_count(void)
 375 {
 376         __this_cpu_inc(hrtimer_interrupts);
 377 }
 378
 379 static int watchdog_nmi_enable(unsigned int cpu);
 380 static void watchdog_nmi_disable(unsigned int cpu);
 381
 382 static int watchdog_enable_all_cpus(void);
 383 static void watchdog_disable_all_cpus(void);
 384
 385 /* watchdog kicker functions */
 386 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 387 {
 388         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 389         struct pt_regs *regs = get_irq_regs();
 390         int duration;
 391         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 392
 393         /* kick the hardlockup detector */
 394         watchdog_interrupt_count();
 395
 396         /* kick the softlockup detector */
 397         wake_up_process(__this_cpu_read(softlockup_watchdog));
 398
 399         /* .. and repeat */
 400         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 401
 402         if (touch_ts == 0) {
 403                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 404                         /*
 405                          * If the time stamp was touched atomically
 406                          * make sure the scheduler tick is up to date.
 407                          */
 408                         __this_cpu_write(softlockup_touch_sync, false);
 409                         sched_clock_tick();
 410                 }
 411
 412                 /* Clear the guest paused flag on watchdog reset */
 413                 kvm_check_and_clear_guest_paused();
 414                 __touch_watchdog();
 415                 return HRTIMER_RESTART;
 416         }
 417
 418         /* check for a softlockup
 419          * This is done by making sure a high priority task is
 420          * being scheduled.  The task touches the watchdog to
 421          * indicate it is getting cpu time.  If it hasn't then
 422          * this is a good indication some task is hogging the cpu
 423          */
 424         duration = is_softlockup(touch_ts);
 425         if (unlikely(duration)) {
 426                 /*
 427                  * If a virtual machine is stopped by the host it can look to
 428                  * the watchdog like a soft lockup, check to see if the host
 429                  * stopped the vm before we issue the warning
 430                  */
 431                 if (kvm_check_and_clear_guest_paused())
 432                         return HRTIMER_RESTART;
 433
 434                 /* only warn once */
 435                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 436                         /*
 437                          * When multiple processes are causing softlockups the
 438                          * softlockup detector only warns on the first one
 439                          * because the code relies on a full quiet cycle to
 440                          * re-arm.  The second process prevents the quiet cycle
 441                          * and never gets reported.  Use task pointers to detect
 442                          * this.
 443                          */
 444                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 445                             current) {
 446                                 __this_cpu_write(soft_watchdog_warn, false);
 447                                 __touch_watchdog();
 448                         }
 449                         return HRTIMER_RESTART;
 450                 }
 451
 452                 if (softlockup_all_cpu_backtrace) {
 453                         /* Prevent multiple soft-lockup reports if one cpu is already
 454                          * engaged in dumping cpu back traces
 455                          */
 456                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 457                                 /* Someone else will report us. Let's give up */
 458                                 __this_cpu_write(soft_watchdog_warn, true);
 459                                 return HRTIMER_RESTART;
 460                         }
 461                 }
 462
 463                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 464                         smp_processor_id(), duration,
 465                         current->comm, task_pid_nr(current));
 466                 __this_cpu_write(softlockup_task_ptr_saved, current);
 467                 print_modules();
 468                 print_irqtrace_events(current);
 469                 if (regs)
 470                         show_regs(regs);
 471                 else
 472                         dump_stack();
 473
 474                 if (softlockup_all_cpu_backtrace) {
 475                         /* Avoid generating two back traces for current
 476                          * given that one is already made above
 477                          */
 478                         trigger_allbutself_cpu_backtrace();
 479
 480                         clear_bit(0, &soft_lockup_nmi_warn);
 481                         /* Barrier to sync with other cpus */
 482                         smp_mb__after_atomic();
 483                 }
 484
 485                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 486                 if (softlockup_panic)
 487                         panic("softlockup: hung tasks");
 488                 __this_cpu_write(soft_watchdog_warn, true);
 489         } else
 490                 __this_cpu_write(soft_watchdog_warn, false);
 491
 492         return HRTIMER_RESTART;
 493 }
 494
 495 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 496 {
 497         struct sched_param param = { .sched_priority = prio };
 498
 499         sched_setscheduler(current, policy, &param);
 500 }
 501
 502 static void watchdog_enable(unsigned int cpu)
 503 {
 504         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 505
 506         /* kick off the timer for the hardlockup detector */
 507         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 508         hrtimer->function = watchdog_timer_fn;
 509         hrtimer->irqsafe = 1;
 510
 511         /* Enable the perf event */
 512         watchdog_nmi_enable(cpu);
 513
 514         /* done here because hrtimer_start can only pin to smp_processor_id() */
 515         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 516                       HRTIMER_MODE_REL_PINNED);
 517
 518         /* initialize timestamp */
 519         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 520         __touch_watchdog();
 521 }
 522
 523 static void watchdog_disable(unsigned int cpu)
 524 {
 525         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 526
 527         watchdog_set_prio(SCHED_NORMAL, 0);
 528         hrtimer_cancel(hrtimer);
 529         /* disable the perf event */
 530         watchdog_nmi_disable(cpu);
 531 }
 532
 533 static void watchdog_cleanup(unsigned int cpu, bool online)
 534 {
 535         watchdog_disable(cpu);
 536 }
 537
 538 static int watchdog_should_run(unsigned int cpu)
 539 {
 540         return __this_cpu_read(hrtimer_interrupts) !=
 541                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 542 }
 543
 544 /*
 545  * The watchdog thread function - touches the timestamp.
 546  *
 547  * It only runs once every sample_period seconds (4 seconds by
 548  * default) to reset the softlockup timestamp. If this gets delayed
 549  * for more than 2*watchdog_thresh seconds then the debug-printout
 550  * triggers in watchdog_timer_fn().
 551  */
 552 static void watchdog(unsigned int cpu)
 553 {
 554         __this_cpu_write(soft_lockup_hrtimer_cnt,
 555                          __this_cpu_read(hrtimer_interrupts));
 556         __touch_watchdog();
 557
 558         /*
 559          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 560          * failure path. Check for failures that can occur asynchronously -
 561          * for example, when CPUs are on-lined - and shut down the hardware
 562          * perf event on each CPU accordingly.
 563          *
 564          * The only non-obvious place this bit can be cleared is through
 565          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 566          * pr_info here would be too noisy as it would result in a message
 567          * every few seconds if the hardlockup was disabled but the softlockup
 568          * enabled.
 569          */
 570         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 571                 watchdog_nmi_disable(cpu);
 572 }
 573
 574 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 575 /*
 576  * People like the simple clean cpu node info on boot.
 577  * Reduce the watchdog noise by only printing messages
 578  * that are different from what cpu0 displayed.
 579  */
 580 static unsigned long cpu0_err;
 581
 582 static int watchdog_nmi_enable(unsigned int cpu)
 583 {
 584         struct perf_event_attr *wd_attr;
 585         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 586
 587         /* nothing to do if the hard lockup detector is disabled */
 588         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 589                 goto out;
 590
 591         /* is it already setup and enabled? */
 592         if (event && event->state > PERF_EVENT_STATE_OFF)
 593                 goto out;
 594
 595         /* it is setup but not enabled */
 596         if (event != NULL)
 597                 goto out_enable;
 598
 599         wd_attr = &wd_hw_attr;
 600         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 601
 602         /* Try to register using hardware perf events */
 603         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 604
 605         /* save cpu0 error for future comparision */
 606         if (cpu == 0 && IS_ERR(event))
 607                 cpu0_err = PTR_ERR(event);
 608
 609         if (!IS_ERR(event)) {
 610                 /* only print for cpu0 or different than cpu0 */
 611                 if (cpu == 0 || cpu0_err)
 612                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 613                 goto out_save;
 614         }
 615
 616         /*
 617          * Disable the hard lockup detector if _any_ CPU fails to set up
 618          * set up the hardware perf event. The watchdog() function checks
 619          * the NMI_WATCHDOG_ENABLED bit periodically.
 620          *
 621          * The barriers are for syncing up watchdog_enabled across all the
 622          * cpus, as clear_bit() does not use barriers.
 623          */
 624         smp_mb__before_atomic();
 625         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 626         smp_mb__after_atomic();
 627
 628         /* skip displaying the same error again */
 629         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 630                 return PTR_ERR(event);
 631
 632         /* vary the KERN level based on the returned errno */
 633         if (PTR_ERR(event) == -EOPNOTSUPP)
 634                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 635         else if (PTR_ERR(event) == -ENOENT)
 636                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 637                          cpu);
 638         else
 639                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 640                         cpu, PTR_ERR(event));
 641
 642         pr_info("Shutting down hard lockup detector on all cpus\n");
 643
 644         return PTR_ERR(event);
 645
 646         /* success path */
 647 out_save:
 648         per_cpu(watchdog_ev, cpu) = event;
 649 out_enable:
 650         perf_event_enable(per_cpu(watchdog_ev, cpu));
 651 out:
 652         return 0;
 653 }
 654
 655 static void watchdog_nmi_disable(unsigned int cpu)
 656 {
 657         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 658
 659         if (event) {
 660                 perf_event_disable(event);
 661                 per_cpu(watchdog_ev, cpu) = NULL;
 662
 663                 /* should be in cleanup, but blocks oprofile */
 664                 perf_event_release_kernel(event);
 665         }
 666         if (cpu == 0) {
 667                 /* watchdog_nmi_enable() expects this to be zero initially. */
 668                 cpu0_err = 0;
 669         }
 670 }
 671
 672 #else
 673 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 674 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 675 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 676
 677 static struct smp_hotplug_thread watchdog_threads = {
 678         .store                  = &softlockup_watchdog,
 679         .thread_should_run      = watchdog_should_run,
 680         .thread_fn              = watchdog,
 681         .thread_comm            = "watchdog/%u",
 682         .setup                  = watchdog_enable,
 683         .cleanup                = watchdog_cleanup,
 684         .park                   = watchdog_disable,
 685         .unpark                 = watchdog_enable,
 686 };
 687
 688 /*
 689  * park all watchdog threads that are specified in 'watchdog_cpumask'
 690  *
 691  * This function returns an error if kthread_park() of a watchdog thread
 692  * fails. In this situation, the watchdog threads of some CPUs can already
 693  * be parked and the watchdog threads of other CPUs can still be runnable.
 694  * Callers are expected to handle this special condition as appropriate in
 695  * their context.
 696  *
 697  * This function may only be called in a context that is protected against
 698  * races with CPU hotplug - for example, via get_online_cpus().
 699  */
 700 static int watchdog_park_threads(void)
 701 {
 702         int cpu, ret = 0;
 703
 704         for_each_watchdog_cpu(cpu) {
 705                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 706                 if (ret)
 707                         break;
 708         }
 709
 710         return ret;
 711 }
 712
 713 /*
 714  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 715  *
 716  * This function may only be called in a context that is protected against
 717  * races with CPU hotplug - for example, via get_online_cpus().
 718  */
 719 static void watchdog_unpark_threads(void)
 720 {
 721         int cpu;
 722
 723         for_each_watchdog_cpu(cpu)
 724                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 725 }
 726
 727 /*
 728  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 729  */
 730 int lockup_detector_suspend(void)
 731 {
 732         int ret = 0;
 733
 734         get_online_cpus();
 735         mutex_lock(&watchdog_proc_mutex);
 736         /*
 737          * Multiple suspend requests can be active in parallel (counted by
 738          * the 'watchdog_suspended' variable). If the watchdog threads are
 739          * running, the first caller takes care that they will be parked.
 740          * The state of 'watchdog_running' cannot change while a suspend
 741          * request is active (see related code in 'proc' handlers).
 742          */
 743         if (watchdog_running && !watchdog_suspended)
 744                 ret = watchdog_park_threads();
 745
 746         if (ret == 0)
 747                 watchdog_suspended++;
 748         else {
 749                 watchdog_disable_all_cpus();
 750                 pr_err("Failed to suspend lockup detectors, disabled\n");
 751                 watchdog_enabled = 0;
 752         }
 753
 754         mutex_unlock(&watchdog_proc_mutex);
 755
 756         return ret;
 757 }
 758
 759 /*
 760  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 761  */
 762 void lockup_detector_resume(void)
 763 {
 764         mutex_lock(&watchdog_proc_mutex);
 765
 766         watchdog_suspended--;
 767         /*
 768          * The watchdog threads are unparked if they were previously running
 769          * and if there is no more active suspend request.
 770          */
 771         if (watchdog_running && !watchdog_suspended)
 772                 watchdog_unpark_threads();
 773
 774         mutex_unlock(&watchdog_proc_mutex);
 775         put_online_cpus();
 776 }
 777
 778 static int update_watchdog_all_cpus(void)
 779 {
 780         int ret;
 781
 782         ret = watchdog_park_threads();
 783         if (ret)
 784                 return ret;
 785
 786         watchdog_unpark_threads();
 787
 788         return 0;
 789 }
 790
 791 static int watchdog_enable_all_cpus(void)
 792 {
 793         int err = 0;
 794
 795         if (!watchdog_running) {
 796                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 797                                                              &watchdog_cpumask);
 798                 if (err)
 799                         pr_err("Failed to create watchdog threads, disabled\n");
 800                 else
 801                         watchdog_running = 1;
 802         } else {
 803                 /*
 804                  * Enable/disable the lockup detectors or
 805                  * change the sample period 'on the fly'.
 806                  */
 807                 err = update_watchdog_all_cpus();
 808
 809                 if (err) {
 810                         watchdog_disable_all_cpus();
 811                         pr_err("Failed to update lockup detectors, disabled\n");
 812                 }
 813         }
 814
 815         if (err)
 816                 watchdog_enabled = 0;
 817
 818         return err;
 819 }
 820
 821 static void watchdog_disable_all_cpus(void)
 822 {
 823         if (watchdog_running) {
 824                 watchdog_running = 0;
 825                 smpboot_unregister_percpu_thread(&watchdog_threads);
 826         }
 827 }
 828
 829 #ifdef CONFIG_SYSCTL
 830
 831 /*
 832  * Update the run state of the lockup detectors.
 833  */
 834 static int proc_watchdog_update(void)
 835 {
 836         int err = 0;
 837
 838         /*
 839          * Watchdog threads won't be started if they are already active.
 840          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 841          * care of this. If those threads are already active, the sample
 842          * period will be updated and the lockup detectors will be enabled
 843          * or disabled 'on the fly'.
 844          */
 845         if (watchdog_enabled && watchdog_thresh)
 846                 err = watchdog_enable_all_cpus();
 847         else
 848                 watchdog_disable_all_cpus();
 849
 850         return err;
 851
 852 }
 853
 854 /*
 855  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 856  *
 857  * caller             | table->data points to | 'which' contains the flag(s)
 858  * -------------------|-----------------------|-----------------------------
 859  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 860  *                    |                       | with SOFT_WATCHDOG_ENABLED
 861  * -------------------|-----------------------|-----------------------------
 862  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 863  * -------------------|-----------------------|-----------------------------
 864  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 865  */
 866 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 867                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 868 {
 869         int err, old, new;
 870         int *watchdog_param = (int *)table->data;
 871
 872         get_online_cpus();
 873         mutex_lock(&watchdog_proc_mutex);
 874
 875         if (watchdog_suspended) {
 876                 /* no parameter changes allowed while watchdog is suspended */
 877                 err = -EAGAIN;
 878                 goto out;
 879         }
 880
 881         /*
 882          * If the parameter is being read return the state of the corresponding
 883          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 884          * run state of the lockup detectors.
 885          */
 886         if (!write) {
 887                 *watchdog_param = (watchdog_enabled & which) != 0;
 888                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 889         } else {
 890                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 891                 if (err)
 892                         goto out;
 893
 894                 /*
 895                  * There is a race window between fetching the current value
 896                  * from 'watchdog_enabled' and storing the new value. During
 897                  * this race window, watchdog_nmi_enable() can sneak in and
 898                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 899                  * The 'cmpxchg' detects this race and the loop retries.
 900                  */
 901                 do {
 902                         old = watchdog_enabled;
 903                         /*
 904                          * If the parameter value is not zero set the
 905                          * corresponding bit(s), else clear it(them).
 906                          */
 907                         if (*watchdog_param)
 908                                 new = old | which;
 909                         else
 910                                 new = old & ~which;
 911                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 912
 913                 /*
 914                  * Update the run state of the lockup detectors. There is _no_
 915                  * need to check the value returned by proc_watchdog_update()
 916                  * and to restore the previous value of 'watchdog_enabled' as
 917                  * both lockup detectors are disabled if proc_watchdog_update()
 918                  * returns an error.
 919                  */
 920                 if (old == new)
 921                         goto out;
 922
 923                 err = proc_watchdog_update();
 924         }
 925 out:
 926         mutex_unlock(&watchdog_proc_mutex);
 927         put_online_cpus();
 928         return err;
 929 }
 930
 931 /*
 932  * /proc/sys/kernel/watchdog
 933  */
 934 int proc_watchdog(struct ctl_table *table, int write,
 935                   void __user *buffer, size_t *lenp, loff_t *ppos)
 936 {
 937         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 938                                     table, write, buffer, lenp, ppos);
 939 }
 940
 941 /*
 942  * /proc/sys/kernel/nmi_watchdog
 943  */
 944 int proc_nmi_watchdog(struct ctl_table *table, int write,
 945                       void __user *buffer, size_t *lenp, loff_t *ppos)
 946 {
 947         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 948                                     table, write, buffer, lenp, ppos);
 949 }
 950
 951 /*
 952  * /proc/sys/kernel/soft_watchdog
 953  */
 954 int proc_soft_watchdog(struct ctl_table *table, int write,
 955                         void __user *buffer, size_t *lenp, loff_t *ppos)
 956 {
 957         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 958                                     table, write, buffer, lenp, ppos);
 959 }
 960
 961 /*
 962  * /proc/sys/kernel/watchdog_thresh
 963  */
 964 int proc_watchdog_thresh(struct ctl_table *table, int write,
 965                          void __user *buffer, size_t *lenp, loff_t *ppos)
 966 {
 967         int err, old, new;
 968
 969         get_online_cpus();
 970         mutex_lock(&watchdog_proc_mutex);
 971
 972         if (watchdog_suspended) {
 973                 /* no parameter changes allowed while watchdog is suspended */
 974                 err = -EAGAIN;
 975                 goto out;
 976         }
 977
 978         old = ACCESS_ONCE(watchdog_thresh);
 979         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 980
 981         if (err || !write)
 982                 goto out;
 983
 984         /*
 985          * Update the sample period. Restore on failure.
 986          */
 987         new = ACCESS_ONCE(watchdog_thresh);
 988         if (old == new)
 989                 goto out;
 990
 991         set_sample_period();
 992         err = proc_watchdog_update();
 993         if (err) {
 994                 watchdog_thresh = old;
 995                 set_sample_period();
 996         }
 997 out:
 998         mutex_unlock(&watchdog_proc_mutex);
 999         put_online_cpus();
1000         return err;
1001 }
1002
1003 /*
1004  * The cpumask is the mask of possible cpus that the watchdog can run
1005  * on, not the mask of cpus it is actually running on.  This allows the
1006  * user to specify a mask that will include cpus that have not yet
1007  * been brought online, if desired.
1008  */
1009 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1010                           void __user *buffer, size_t *lenp, loff_t *ppos)
1011 {
1012         int err;
1013
1014         get_online_cpus();
1015         mutex_lock(&watchdog_proc_mutex);
1016
1017         if (watchdog_suspended) {
1018                 /* no parameter changes allowed while watchdog is suspended */
1019                 err = -EAGAIN;
1020                 goto out;
1021         }
1022
1023         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1024         if (!err && write) {
1025                 /* Remove impossible cpus to keep sysctl output cleaner. */
1026                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1027                             cpu_possible_mask);
1028
1029                 if (watchdog_running) {
1030                         /*
1031                          * Failure would be due to being unable to allocate
1032                          * a temporary cpumask, so we are likely not in a
1033                          * position to do much else to make things better.
1034                          */
1035                         if (smpboot_update_cpumask_percpu_thread(
1036                                     &watchdog_threads, &watchdog_cpumask) != 0)
1037                                 pr_err("cpumask update failed\n");
1038                 }
1039         }
1040 out:
1041         mutex_unlock(&watchdog_proc_mutex);
1042         put_online_cpus();
1043         return err;
1044 }
1045
1046 #endif /* CONFIG_SYSCTL */
1047
1048 void __init lockup_detector_init(void)
1049 {
1050         set_sample_period();
1051
1052 #ifdef CONFIG_NO_HZ_FULL
1053         if (tick_nohz_full_enabled()) {
1054                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1055                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1056         } else
1057                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1058 #else
1059         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1060 #endif
1061
1062         if (watchdog_enabled)
1063                 watchdog_enable_all_cpus();
1064 }