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 #include <linux/workqueue.h>
  24
  25 #include <asm/irq_regs.h>
  26 #include <linux/kvm_para.h>
  27 #include <linux/perf_event.h>
  28 #include <linux/kthread.h>
  29
  30 /*
  31  * The run state of the lockup detectors is controlled by the content of the
  32  * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
  33  * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
  34  *
  35  * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
  36  * are variables that are only used as an 'interface' between the parameters
  37  * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
  38  * 'watchdog_thresh' variable is handled differently because its value is not
  39  * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
  40  * is equal zero.
  41  */
  42 #define NMI_WATCHDOG_ENABLED_BIT   0
  43 #define SOFT_WATCHDOG_ENABLED_BIT  1
  44 #define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
  45 #define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
  46
  47 static DEFINE_MUTEX(watchdog_proc_mutex);
  48
  49 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  50 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
  51 #else
  52 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  53 #endif
  54 int __read_mostly nmi_watchdog_enabled;
  55 int __read_mostly soft_watchdog_enabled;
  56 int __read_mostly watchdog_user_enabled;
  57 int __read_mostly watchdog_thresh = 10;
  58
  59 #ifdef CONFIG_SMP
  60 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  61 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  62 #else
  63 #define sysctl_softlockup_all_cpu_backtrace 0
  64 #define sysctl_hardlockup_all_cpu_backtrace 0
  65 #endif
  66 static struct cpumask watchdog_cpumask __read_mostly;
  67 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  68
  69 /* Helper for online, unparked cpus. */
  70 #define for_each_watchdog_cpu(cpu) \
  71         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
  72
  73 /*
  74  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  75  * are registered/started and is set to 0 when the watchdog threads are
  76  * unregistered/stopped, so it is an indicator whether the threads exist.
  77  */
  78 static int __read_mostly watchdog_running;
  79 /*
  80  * If a subsystem has a need to deactivate the watchdog temporarily, it
  81  * can use the suspend/resume interface to achieve this. The content of
  82  * the 'watchdog_suspended' variable reflects this state. Existing threads
  83  * are parked/unparked by the lockup_detector_{suspend|resume} functions
  84  * (see comment blocks pertaining to those functions for further details).
  85  *
  86  * 'watchdog_suspended' also prevents threads from being registered/started
  87  * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
  88  * of 'watchdog_running' cannot change while the watchdog is deactivated
  89  * temporarily (see related code in 'proc' handlers).
  90  */
  91 static int __read_mostly watchdog_suspended;
  92
  93 static u64 __read_mostly sample_period;
  94
  95 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  96 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  97 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  98 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  99 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 100 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 101 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 102 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 103 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 104 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
 105 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
 106 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 107 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
 108 #endif
 109 static unsigned long soft_lockup_nmi_warn;
 110
 111 /* boot commands */
 112 /*
 113  * Should we panic when a soft-lockup or hard-lockup occurs:
 114  */
 115 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 116 unsigned int __read_mostly hardlockup_panic =
 117                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
 118 static unsigned long hardlockup_allcpu_dumped;
 119 /*
 120  * We may not want to enable hard lockup detection by default in all cases,
 121  * for example when running the kernel as a guest on a hypervisor. In these
 122  * cases this function can be called to disable hard lockup detection. This
 123  * function should only be executed once by the boot processor before the
 124  * kernel command line parameters are parsed, because otherwise it is not
 125  * possible to override this in hardlockup_panic_setup().
 126  */
 127 void hardlockup_detector_disable(void)
 128 {
 129         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 130 }
 131
 132 static int __init hardlockup_panic_setup(char *str)
 133 {
 134         if (!strncmp(str, "panic", 5))
 135                 hardlockup_panic = 1;
 136         else if (!strncmp(str, "nopanic", 7))
 137                 hardlockup_panic = 0;
 138         else if (!strncmp(str, "0", 1))
 139                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 140         else if (!strncmp(str, "1", 1))
 141                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 142         return 1;
 143 }
 144 __setup("nmi_watchdog=", hardlockup_panic_setup);
 145 #endif
 146
 147 unsigned int __read_mostly softlockup_panic =
 148                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 149
 150 static int __init softlockup_panic_setup(char *str)
 151 {
 152         softlockup_panic = simple_strtoul(str, NULL, 0);
 153
 154         return 1;
 155 }
 156 __setup("softlockup_panic=", softlockup_panic_setup);
 157
 158 static int __init nowatchdog_setup(char *str)
 159 {
 160         watchdog_enabled = 0;
 161         return 1;
 162 }
 163 __setup("nowatchdog", nowatchdog_setup);
 164
 165 static int __init nosoftlockup_setup(char *str)
 166 {
 167         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 168         return 1;
 169 }
 170 __setup("nosoftlockup", nosoftlockup_setup);
 171
 172 #ifdef CONFIG_SMP
 173 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 174 {
 175         sysctl_softlockup_all_cpu_backtrace =
 176                 !!simple_strtol(str, NULL, 0);
 177         return 1;
 178 }
 179 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 180 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
 181 {
 182         sysctl_hardlockup_all_cpu_backtrace =
 183                 !!simple_strtol(str, NULL, 0);
 184         return 1;
 185 }
 186 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
 187 #endif
 188
 189 /*
 190  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 191  * lockups can have false positives under extreme conditions. So we generally
 192  * want a higher threshold for soft lockups than for hard lockups. So we couple
 193  * the thresholds with a factor: we make the soft threshold twice the amount of
 194  * time the hard threshold is.
 195  */
 196 static int get_softlockup_thresh(void)
 197 {
 198         return watchdog_thresh * 2;
 199 }
 200
 201 /*
 202  * Returns seconds, approximately.  We don't need nanosecond
 203  * resolution, and we don't need to waste time with a big divide when
 204  * 2^30ns == 1.074s.
 205  */
 206 static unsigned long get_timestamp(void)
 207 {
 208         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 209 }
 210
 211 static void set_sample_period(void)
 212 {
 213         /*
 214          * convert watchdog_thresh from seconds to ns
 215          * the divide by 5 is to give hrtimer several chances (two
 216          * or three with the current relation between the soft
 217          * and hard thresholds) to increment before the
 218          * hardlockup detector generates a warning
 219          */
 220         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 221 }
 222
 223 /* Commands for resetting the watchdog */
 224 static void __touch_watchdog(void)
 225 {
 226         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 227 }
 228
 229 /**
 230  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 231  *
 232  * Call when the scheduler may have stalled for legitimate reasons
 233  * preventing the watchdog task from executing - e.g. the scheduler
 234  * entering idle state.  This should only be used for scheduler events.
 235  * Use touch_softlockup_watchdog() for everything else.
 236  */
 237 void touch_softlockup_watchdog_sched(void)
 238 {
 239         /*
 240          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 241          * gets zeroed here, so use the raw_ operation.
 242          */
 243         raw_cpu_write(watchdog_touch_ts, 0);
 244 }
 245
 246 void touch_softlockup_watchdog(void)
 247 {
 248         touch_softlockup_watchdog_sched();
 249         wq_watchdog_touch(raw_smp_processor_id());
 250 }
 251 EXPORT_SYMBOL(touch_softlockup_watchdog);
 252
 253 void touch_all_softlockup_watchdogs(void)
 254 {
 255         int cpu;
 256
 257         /*
 258          * this is done lockless
 259          * do we care if a 0 races with a timestamp?
 260          * all it means is the softlock check starts one cycle later
 261          */
 262         for_each_watchdog_cpu(cpu)
 263                 per_cpu(watchdog_touch_ts, cpu) = 0;
 264         wq_watchdog_touch(-1);
 265 }
 266
 267 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 268 void touch_nmi_watchdog(void)
 269 {
 270         /*
 271          * Using __raw here because some code paths have
 272          * preemption enabled.  If preemption is enabled
 273          * then interrupts should be enabled too, in which
 274          * case we shouldn't have to worry about the watchdog
 275          * going off.
 276          */
 277         raw_cpu_write(watchdog_nmi_touch, true);
 278         touch_softlockup_watchdog();
 279 }
 280 EXPORT_SYMBOL(touch_nmi_watchdog);
 281
 282 #endif
 283
 284 void touch_softlockup_watchdog_sync(void)
 285 {
 286         __this_cpu_write(softlockup_touch_sync, true);
 287         __this_cpu_write(watchdog_touch_ts, 0);
 288 }
 289
 290 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 291 /* watchdog detector functions */
 292 static bool is_hardlockup(void)
 293 {
 294         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 295
 296         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 297                 return true;
 298
 299         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 300         return false;
 301 }
 302 #endif
 303
 304 static int is_softlockup(unsigned long touch_ts)
 305 {
 306         unsigned long now = get_timestamp();
 307
 308         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 309                 /* Warn about unreasonable delays. */
 310                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 311                         return now - touch_ts;
 312         }
 313         return 0;
 314 }
 315
 316 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 317
 318 /* Can be overriden by architecture */
 319 __weak int hw_nmi_get_event(void)
 320 {
 321         return PERF_COUNT_HW_CPU_CYCLES;
 322 }
 323
 324 static struct perf_event_attr wd_hw_attr = {
 325         .type           = PERF_TYPE_HARDWARE,
 326         .config         = PERF_COUNT_HW_CPU_CYCLES,
 327         .size           = sizeof(struct perf_event_attr),
 328         .pinned         = 1,
 329         .disabled       = 1,
 330 };
 331
 332 /* Callback function for perf event subsystem */
 333 static void watchdog_overflow_callback(struct perf_event *event,
 334                  struct perf_sample_data *data,
 335                  struct pt_regs *regs)
 336 {
 337         /* Ensure the watchdog never gets throttled */
 338         event->hw.interrupts = 0;
 339
 340         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 341                 __this_cpu_write(watchdog_nmi_touch, false);
 342                 return;
 343         }
 344
 345         /* check for a hardlockup
 346          * This is done by making sure our timer interrupt
 347          * is incrementing.  The timer interrupt should have
 348          * fired multiple times before we overflow'd.  If it hasn't
 349          * then this is a good indication the cpu is stuck
 350          */
 351         if (is_hardlockup()) {
 352                 int this_cpu = smp_processor_id();
 353                 struct pt_regs *regs = get_irq_regs();
 354
 355                 /* only print hardlockups once */
 356                 if (__this_cpu_read(hard_watchdog_warn) == true)
 357                         return;
 358
 359                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 360                 print_modules();
 361                 print_irqtrace_events(current);
 362                 if (regs)
 363                         show_regs(regs);
 364                 else
 365                         dump_stack();
 366
 367                 /*
 368                  * Perform all-CPU dump only once to avoid multiple hardlockups
 369                  * generating interleaving traces
 370                  */
 371                 if (sysctl_hardlockup_all_cpu_backtrace &&
 372                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 373                         trigger_allbutself_cpu_backtrace();
 374
 375                 if (hardlockup_panic)
 376                         nmi_panic(regs, "Hard LOCKUP");
 377
 378                 __this_cpu_write(hard_watchdog_warn, true);
 379                 return;
 380         }
 381
 382         __this_cpu_write(hard_watchdog_warn, false);
 383         return;
 384 }
 385 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 386
 387 static void watchdog_interrupt_count(void)
 388 {
 389         __this_cpu_inc(hrtimer_interrupts);
 390 }
 391
 392 static int watchdog_nmi_enable(unsigned int cpu);
 393 static void watchdog_nmi_disable(unsigned int cpu);
 394
 395 static int watchdog_enable_all_cpus(void);
 396 static void watchdog_disable_all_cpus(void);
 397
 398 /* watchdog kicker functions */
 399 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 400 {
 401         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 402         struct pt_regs *regs = get_irq_regs();
 403         int duration;
 404         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 405
 406         /* kick the hardlockup detector */
 407         watchdog_interrupt_count();
 408
 409         /* kick the softlockup detector */
 410         wake_up_process(__this_cpu_read(softlockup_watchdog));
 411
 412         /* .. and repeat */
 413         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 414
 415         if (touch_ts == 0) {
 416                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 417                         /*
 418                          * If the time stamp was touched atomically
 419                          * make sure the scheduler tick is up to date.
 420                          */
 421                         __this_cpu_write(softlockup_touch_sync, false);
 422                         sched_clock_tick();
 423                 }
 424
 425                 /* Clear the guest paused flag on watchdog reset */
 426                 kvm_check_and_clear_guest_paused();
 427                 __touch_watchdog();
 428                 return HRTIMER_RESTART;
 429         }
 430
 431         /* check for a softlockup
 432          * This is done by making sure a high priority task is
 433          * being scheduled.  The task touches the watchdog to
 434          * indicate it is getting cpu time.  If it hasn't then
 435          * this is a good indication some task is hogging the cpu
 436          */
 437         duration = is_softlockup(touch_ts);
 438         if (unlikely(duration)) {
 439                 /*
 440                  * If a virtual machine is stopped by the host it can look to
 441                  * the watchdog like a soft lockup, check to see if the host
 442                  * stopped the vm before we issue the warning
 443                  */
 444                 if (kvm_check_and_clear_guest_paused())
 445                         return HRTIMER_RESTART;
 446
 447                 /* only warn once */
 448                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 449                         /*
 450                          * When multiple processes are causing softlockups the
 451                          * softlockup detector only warns on the first one
 452                          * because the code relies on a full quiet cycle to
 453                          * re-arm.  The second process prevents the quiet cycle
 454                          * and never gets reported.  Use task pointers to detect
 455                          * this.
 456                          */
 457                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 458                             current) {
 459                                 __this_cpu_write(soft_watchdog_warn, false);
 460                                 __touch_watchdog();
 461                         }
 462                         return HRTIMER_RESTART;
 463                 }
 464
 465                 if (softlockup_all_cpu_backtrace) {
 466                         /* Prevent multiple soft-lockup reports if one cpu is already
 467                          * engaged in dumping cpu back traces
 468                          */
 469                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 470                                 /* Someone else will report us. Let's give up */
 471                                 __this_cpu_write(soft_watchdog_warn, true);
 472                                 return HRTIMER_RESTART;
 473                         }
 474                 }
 475
 476                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 477                         smp_processor_id(), duration,
 478                         current->comm, task_pid_nr(current));
 479                 __this_cpu_write(softlockup_task_ptr_saved, current);
 480                 print_modules();
 481                 print_irqtrace_events(current);
 482                 if (regs)
 483                         show_regs(regs);
 484                 else
 485                         dump_stack();
 486
 487                 if (softlockup_all_cpu_backtrace) {
 488                         /* Avoid generating two back traces for current
 489                          * given that one is already made above
 490                          */
 491                         trigger_allbutself_cpu_backtrace();
 492
 493                         clear_bit(0, &soft_lockup_nmi_warn);
 494                         /* Barrier to sync with other cpus */
 495                         smp_mb__after_atomic();
 496                 }
 497
 498                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 499                 if (softlockup_panic)
 500                         panic("softlockup: hung tasks");
 501                 __this_cpu_write(soft_watchdog_warn, true);
 502         } else
 503                 __this_cpu_write(soft_watchdog_warn, false);
 504
 505         return HRTIMER_RESTART;
 506 }
 507
 508 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 509 {
 510         struct sched_param param = { .sched_priority = prio };
 511
 512         sched_setscheduler(current, policy, &param);
 513 }
 514
 515 static void watchdog_enable(unsigned int cpu)
 516 {
 517         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 518
 519         /* kick off the timer for the hardlockup detector */
 520         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 521         hrtimer->function = watchdog_timer_fn;
 522
 523         /* Enable the perf event */
 524         watchdog_nmi_enable(cpu);
 525
 526         /* done here because hrtimer_start can only pin to smp_processor_id() */
 527         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 528                       HRTIMER_MODE_REL_PINNED);
 529
 530         /* initialize timestamp */
 531         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 532         __touch_watchdog();
 533 }
 534
 535 static void watchdog_disable(unsigned int cpu)
 536 {
 537         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 538
 539         watchdog_set_prio(SCHED_NORMAL, 0);
 540         hrtimer_cancel(hrtimer);
 541         /* disable the perf event */
 542         watchdog_nmi_disable(cpu);
 543 }
 544
 545 static void watchdog_cleanup(unsigned int cpu, bool online)
 546 {
 547         watchdog_disable(cpu);
 548 }
 549
 550 static int watchdog_should_run(unsigned int cpu)
 551 {
 552         return __this_cpu_read(hrtimer_interrupts) !=
 553                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 554 }
 555
 556 /*
 557  * The watchdog thread function - touches the timestamp.
 558  *
 559  * It only runs once every sample_period seconds (4 seconds by
 560  * default) to reset the softlockup timestamp. If this gets delayed
 561  * for more than 2*watchdog_thresh seconds then the debug-printout
 562  * triggers in watchdog_timer_fn().
 563  */
 564 static void watchdog(unsigned int cpu)
 565 {
 566         __this_cpu_write(soft_lockup_hrtimer_cnt,
 567                          __this_cpu_read(hrtimer_interrupts));
 568         __touch_watchdog();
 569
 570         /*
 571          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 572          * failure path. Check for failures that can occur asynchronously -
 573          * for example, when CPUs are on-lined - and shut down the hardware
 574          * perf event on each CPU accordingly.
 575          *
 576          * The only non-obvious place this bit can be cleared is through
 577          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 578          * pr_info here would be too noisy as it would result in a message
 579          * every few seconds if the hardlockup was disabled but the softlockup
 580          * enabled.
 581          */
 582         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 583                 watchdog_nmi_disable(cpu);
 584 }
 585
 586 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 587 /*
 588  * People like the simple clean cpu node info on boot.
 589  * Reduce the watchdog noise by only printing messages
 590  * that are different from what cpu0 displayed.
 591  */
 592 static unsigned long cpu0_err;
 593
 594 static int watchdog_nmi_enable(unsigned int cpu)
 595 {
 596         struct perf_event_attr *wd_attr;
 597         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 598
 599         /* nothing to do if the hard lockup detector is disabled */
 600         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 601                 goto out;
 602
 603         /* is it already setup and enabled? */
 604         if (event && event->state > PERF_EVENT_STATE_OFF)
 605                 goto out;
 606
 607         /* it is setup but not enabled */
 608         if (event != NULL)
 609                 goto out_enable;
 610
 611         wd_attr = &wd_hw_attr;
 612         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 613         wd_attr->config = hw_nmi_get_event();
 614
 615         /* Try to register using hardware perf events */
 616         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 617
 618         /* save cpu0 error for future comparision */
 619         if (cpu == 0 && IS_ERR(event))
 620                 cpu0_err = PTR_ERR(event);
 621
 622         if (!IS_ERR(event)) {
 623                 /* only print for cpu0 or different than cpu0 */
 624                 if (cpu == 0 || cpu0_err)
 625                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 626                 goto out_save;
 627         }
 628
 629         /*
 630          * Disable the hard lockup detector if _any_ CPU fails to set up
 631          * set up the hardware perf event. The watchdog() function checks
 632          * the NMI_WATCHDOG_ENABLED bit periodically.
 633          *
 634          * The barriers are for syncing up watchdog_enabled across all the
 635          * cpus, as clear_bit() does not use barriers.
 636          */
 637         smp_mb__before_atomic();
 638         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 639         smp_mb__after_atomic();
 640
 641         /* skip displaying the same error again */
 642         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 643                 return PTR_ERR(event);
 644
 645         /* vary the KERN level based on the returned errno */
 646         if (PTR_ERR(event) == -EOPNOTSUPP)
 647                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 648         else if (PTR_ERR(event) == -ENOENT)
 649                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 650                          cpu);
 651         else
 652                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 653                         cpu, PTR_ERR(event));
 654
 655         pr_info("Shutting down hard lockup detector on all cpus\n");
 656
 657         return PTR_ERR(event);
 658
 659         /* success path */
 660 out_save:
 661         per_cpu(watchdog_ev, cpu) = event;
 662 out_enable:
 663         perf_event_enable(per_cpu(watchdog_ev, cpu));
 664 out:
 665         return 0;
 666 }
 667
 668 static void watchdog_nmi_disable(unsigned int cpu)
 669 {
 670         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 671
 672         if (event) {
 673                 perf_event_disable(event);
 674                 per_cpu(watchdog_ev, cpu) = NULL;
 675
 676                 /* should be in cleanup, but blocks oprofile */
 677                 perf_event_release_kernel(event);
 678         }
 679         if (cpu == 0) {
 680                 /* watchdog_nmi_enable() expects this to be zero initially. */
 681                 cpu0_err = 0;
 682         }
 683 }
 684
 685 #else
 686 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 687 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 688 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 689
 690 static struct smp_hotplug_thread watchdog_threads = {
 691         .store                  = &softlockup_watchdog,
 692         .thread_should_run      = watchdog_should_run,
 693         .thread_fn              = watchdog,
 694         .thread_comm            = "watchdog/%u",
 695         .setup                  = watchdog_enable,
 696         .cleanup                = watchdog_cleanup,
 697         .park                   = watchdog_disable,
 698         .unpark                 = watchdog_enable,
 699 };
 700
 701 /*
 702  * park all watchdog threads that are specified in 'watchdog_cpumask'
 703  *
 704  * This function returns an error if kthread_park() of a watchdog thread
 705  * fails. In this situation, the watchdog threads of some CPUs can already
 706  * be parked and the watchdog threads of other CPUs can still be runnable.
 707  * Callers are expected to handle this special condition as appropriate in
 708  * their context.
 709  *
 710  * This function may only be called in a context that is protected against
 711  * races with CPU hotplug - for example, via get_online_cpus().
 712  */
 713 static int watchdog_park_threads(void)
 714 {
 715         int cpu, ret = 0;
 716
 717         for_each_watchdog_cpu(cpu) {
 718                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 719                 if (ret)
 720                         break;
 721         }
 722
 723         return ret;
 724 }
 725
 726 /*
 727  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 728  *
 729  * This function may only be called in a context that is protected against
 730  * races with CPU hotplug - for example, via get_online_cpus().
 731  */
 732 static void watchdog_unpark_threads(void)
 733 {
 734         int cpu;
 735
 736         for_each_watchdog_cpu(cpu)
 737                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 738 }
 739
 740 /*
 741  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 742  */
 743 int lockup_detector_suspend(void)
 744 {
 745         int ret = 0;
 746
 747         get_online_cpus();
 748         mutex_lock(&watchdog_proc_mutex);
 749         /*
 750          * Multiple suspend requests can be active in parallel (counted by
 751          * the 'watchdog_suspended' variable). If the watchdog threads are
 752          * running, the first caller takes care that they will be parked.
 753          * The state of 'watchdog_running' cannot change while a suspend
 754          * request is active (see related code in 'proc' handlers).
 755          */
 756         if (watchdog_running && !watchdog_suspended)
 757                 ret = watchdog_park_threads();
 758
 759         if (ret == 0)
 760                 watchdog_suspended++;
 761         else {
 762                 watchdog_disable_all_cpus();
 763                 pr_err("Failed to suspend lockup detectors, disabled\n");
 764                 watchdog_enabled = 0;
 765         }
 766
 767         mutex_unlock(&watchdog_proc_mutex);
 768
 769         return ret;
 770 }
 771
 772 /*
 773  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 774  */
 775 void lockup_detector_resume(void)
 776 {
 777         mutex_lock(&watchdog_proc_mutex);
 778
 779         watchdog_suspended--;
 780         /*
 781          * The watchdog threads are unparked if they were previously running
 782          * and if there is no more active suspend request.
 783          */
 784         if (watchdog_running && !watchdog_suspended)
 785                 watchdog_unpark_threads();
 786
 787         mutex_unlock(&watchdog_proc_mutex);
 788         put_online_cpus();
 789 }
 790
 791 static int update_watchdog_all_cpus(void)
 792 {
 793         int ret;
 794
 795         ret = watchdog_park_threads();
 796         if (ret)
 797                 return ret;
 798
 799         watchdog_unpark_threads();
 800
 801         return 0;
 802 }
 803
 804 static int watchdog_enable_all_cpus(void)
 805 {
 806         int err = 0;
 807
 808         if (!watchdog_running) {
 809                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 810                                                              &watchdog_cpumask);
 811                 if (err)
 812                         pr_err("Failed to create watchdog threads, disabled\n");
 813                 else
 814                         watchdog_running = 1;
 815         } else {
 816                 /*
 817                  * Enable/disable the lockup detectors or
 818                  * change the sample period 'on the fly'.
 819                  */
 820                 err = update_watchdog_all_cpus();
 821
 822                 if (err) {
 823                         watchdog_disable_all_cpus();
 824                         pr_err("Failed to update lockup detectors, disabled\n");
 825                 }
 826         }
 827
 828         if (err)
 829                 watchdog_enabled = 0;
 830
 831         return err;
 832 }
 833
 834 static void watchdog_disable_all_cpus(void)
 835 {
 836         if (watchdog_running) {
 837                 watchdog_running = 0;
 838                 smpboot_unregister_percpu_thread(&watchdog_threads);
 839         }
 840 }
 841
 842 #ifdef CONFIG_SYSCTL
 843
 844 /*
 845  * Update the run state of the lockup detectors.
 846  */
 847 static int proc_watchdog_update(void)
 848 {
 849         int err = 0;
 850
 851         /*
 852          * Watchdog threads won't be started if they are already active.
 853          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 854          * care of this. If those threads are already active, the sample
 855          * period will be updated and the lockup detectors will be enabled
 856          * or disabled 'on the fly'.
 857          */
 858         if (watchdog_enabled && watchdog_thresh)
 859                 err = watchdog_enable_all_cpus();
 860         else
 861                 watchdog_disable_all_cpus();
 862
 863         return err;
 864
 865 }
 866
 867 /*
 868  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 869  *
 870  * caller             | table->data points to | 'which' contains the flag(s)
 871  * -------------------|-----------------------|-----------------------------
 872  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 873  *                    |                       | with SOFT_WATCHDOG_ENABLED
 874  * -------------------|-----------------------|-----------------------------
 875  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 876  * -------------------|-----------------------|-----------------------------
 877  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 878  */
 879 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 880                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 881 {
 882         int err, old, new;
 883         int *watchdog_param = (int *)table->data;
 884
 885         get_online_cpus();
 886         mutex_lock(&watchdog_proc_mutex);
 887
 888         if (watchdog_suspended) {
 889                 /* no parameter changes allowed while watchdog is suspended */
 890                 err = -EAGAIN;
 891                 goto out;
 892         }
 893
 894         /*
 895          * If the parameter is being read return the state of the corresponding
 896          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 897          * run state of the lockup detectors.
 898          */
 899         if (!write) {
 900                 *watchdog_param = (watchdog_enabled & which) != 0;
 901                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 902         } else {
 903                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 904                 if (err)
 905                         goto out;
 906
 907                 /*
 908                  * There is a race window between fetching the current value
 909                  * from 'watchdog_enabled' and storing the new value. During
 910                  * this race window, watchdog_nmi_enable() can sneak in and
 911                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 912                  * The 'cmpxchg' detects this race and the loop retries.
 913                  */
 914                 do {
 915                         old = watchdog_enabled;
 916                         /*
 917                          * If the parameter value is not zero set the
 918                          * corresponding bit(s), else clear it(them).
 919                          */
 920                         if (*watchdog_param)
 921                                 new = old | which;
 922                         else
 923                                 new = old & ~which;
 924                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 925
 926                 /*
 927                  * Update the run state of the lockup detectors. There is _no_
 928                  * need to check the value returned by proc_watchdog_update()
 929                  * and to restore the previous value of 'watchdog_enabled' as
 930                  * both lockup detectors are disabled if proc_watchdog_update()
 931                  * returns an error.
 932                  */
 933                 if (old == new)
 934                         goto out;
 935
 936                 err = proc_watchdog_update();
 937         }
 938 out:
 939         mutex_unlock(&watchdog_proc_mutex);
 940         put_online_cpus();
 941         return err;
 942 }
 943
 944 /*
 945  * /proc/sys/kernel/watchdog
 946  */
 947 int proc_watchdog(struct ctl_table *table, int write,
 948                   void __user *buffer, size_t *lenp, loff_t *ppos)
 949 {
 950         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 951                                     table, write, buffer, lenp, ppos);
 952 }
 953
 954 /*
 955  * /proc/sys/kernel/nmi_watchdog
 956  */
 957 int proc_nmi_watchdog(struct ctl_table *table, int write,
 958                       void __user *buffer, size_t *lenp, loff_t *ppos)
 959 {
 960         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 961                                     table, write, buffer, lenp, ppos);
 962 }
 963
 964 /*
 965  * /proc/sys/kernel/soft_watchdog
 966  */
 967 int proc_soft_watchdog(struct ctl_table *table, int write,
 968                         void __user *buffer, size_t *lenp, loff_t *ppos)
 969 {
 970         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 971                                     table, write, buffer, lenp, ppos);
 972 }
 973
 974 /*
 975  * /proc/sys/kernel/watchdog_thresh
 976  */
 977 int proc_watchdog_thresh(struct ctl_table *table, int write,
 978                          void __user *buffer, size_t *lenp, loff_t *ppos)
 979 {
 980         int err, old, new;
 981
 982         get_online_cpus();
 983         mutex_lock(&watchdog_proc_mutex);
 984
 985         if (watchdog_suspended) {
 986                 /* no parameter changes allowed while watchdog is suspended */
 987                 err = -EAGAIN;
 988                 goto out;
 989         }
 990
 991         old = ACCESS_ONCE(watchdog_thresh);
 992         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 993
 994         if (err || !write)
 995                 goto out;
 996
 997         /*
 998          * Update the sample period. Restore on failure.
 999          */
1000         new = ACCESS_ONCE(watchdog_thresh);
1001         if (old == new)
1002                 goto out;
1003
1004         set_sample_period();
1005         err = proc_watchdog_update();
1006         if (err) {
1007                 watchdog_thresh = old;
1008                 set_sample_period();
1009         }
1010 out:
1011         mutex_unlock(&watchdog_proc_mutex);
1012         put_online_cpus();
1013         return err;
1014 }
1015
1016 /*
1017  * The cpumask is the mask of possible cpus that the watchdog can run
1018  * on, not the mask of cpus it is actually running on.  This allows the
1019  * user to specify a mask that will include cpus that have not yet
1020  * been brought online, if desired.
1021  */
1022 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1023                           void __user *buffer, size_t *lenp, loff_t *ppos)
1024 {
1025         int err;
1026
1027         get_online_cpus();
1028         mutex_lock(&watchdog_proc_mutex);
1029
1030         if (watchdog_suspended) {
1031                 /* no parameter changes allowed while watchdog is suspended */
1032                 err = -EAGAIN;
1033                 goto out;
1034         }
1035
1036         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1037         if (!err && write) {
1038                 /* Remove impossible cpus to keep sysctl output cleaner. */
1039                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1040                             cpu_possible_mask);
1041
1042                 if (watchdog_running) {
1043                         /*
1044                          * Failure would be due to being unable to allocate
1045                          * a temporary cpumask, so we are likely not in a
1046                          * position to do much else to make things better.
1047                          */
1048                         if (smpboot_update_cpumask_percpu_thread(
1049                                     &watchdog_threads, &watchdog_cpumask) != 0)
1050                                 pr_err("cpumask update failed\n");
1051                 }
1052         }
1053 out:
1054         mutex_unlock(&watchdog_proc_mutex);
1055         put_online_cpus();
1056         return err;
1057 }
1058
1059 #endif /* CONFIG_SYSCTL */
1060
1061 void __init lockup_detector_init(void)
1062 {
1063         set_sample_period();
1064
1065 #ifdef CONFIG_NO_HZ_FULL
1066         if (tick_nohz_full_enabled()) {
1067                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1068                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1069         } else
1070                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1071 #else
1072         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1073 #endif
1074
1075         if (watchdog_enabled)
1076                 watchdog_enable_all_cpus();
1077 }