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 static struct perf_event_attr wd_hw_attr = {
 319         .type           = PERF_TYPE_HARDWARE,
 320         .config         = PERF_COUNT_HW_CPU_CYCLES,
 321         .size           = sizeof(struct perf_event_attr),
 322         .pinned         = 1,
 323         .disabled       = 1,
 324 };
 325
 326 /* Callback function for perf event subsystem */
 327 static void watchdog_overflow_callback(struct perf_event *event,
 328                  struct perf_sample_data *data,
 329                  struct pt_regs *regs)
 330 {
 331         /* Ensure the watchdog never gets throttled */
 332         event->hw.interrupts = 0;
 333
 334         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 335                 __this_cpu_write(watchdog_nmi_touch, false);
 336                 return;
 337         }
 338
 339         /* check for a hardlockup
 340          * This is done by making sure our timer interrupt
 341          * is incrementing.  The timer interrupt should have
 342          * fired multiple times before we overflow'd.  If it hasn't
 343          * then this is a good indication the cpu is stuck
 344          */
 345         if (is_hardlockup()) {
 346                 int this_cpu = smp_processor_id();
 347                 struct pt_regs *regs = get_irq_regs();
 348
 349                 /* only print hardlockups once */
 350                 if (__this_cpu_read(hard_watchdog_warn) == true)
 351                         return;
 352
 353                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 354                 print_modules();
 355                 print_irqtrace_events(current);
 356                 if (regs)
 357                         show_regs(regs);
 358                 else
 359                         dump_stack();
 360
 361                 /*
 362                  * Perform all-CPU dump only once to avoid multiple hardlockups
 363                  * generating interleaving traces
 364                  */
 365                 if (sysctl_hardlockup_all_cpu_backtrace &&
 366                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 367                         trigger_allbutself_cpu_backtrace();
 368
 369                 if (hardlockup_panic)
 370                         nmi_panic(regs, "Hard LOCKUP");
 371
 372                 __this_cpu_write(hard_watchdog_warn, true);
 373                 return;
 374         }
 375
 376         __this_cpu_write(hard_watchdog_warn, false);
 377         return;
 378 }
 379 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 380
 381 static void watchdog_interrupt_count(void)
 382 {
 383         __this_cpu_inc(hrtimer_interrupts);
 384 }
 385
 386 static int watchdog_nmi_enable(unsigned int cpu);
 387 static void watchdog_nmi_disable(unsigned int cpu);
 388
 389 static int watchdog_enable_all_cpus(void);
 390 static void watchdog_disable_all_cpus(void);
 391
 392 /* watchdog kicker functions */
 393 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 394 {
 395         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 396         struct pt_regs *regs = get_irq_regs();
 397         int duration;
 398         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 399
 400         /* kick the hardlockup detector */
 401         watchdog_interrupt_count();
 402
 403         /* kick the softlockup detector */
 404         wake_up_process(__this_cpu_read(softlockup_watchdog));
 405
 406         /* .. and repeat */
 407         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 408
 409         if (touch_ts == 0) {
 410                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 411                         /*
 412                          * If the time stamp was touched atomically
 413                          * make sure the scheduler tick is up to date.
 414                          */
 415                         __this_cpu_write(softlockup_touch_sync, false);
 416                         sched_clock_tick();
 417                 }
 418
 419                 /* Clear the guest paused flag on watchdog reset */
 420                 kvm_check_and_clear_guest_paused();
 421                 __touch_watchdog();
 422                 return HRTIMER_RESTART;
 423         }
 424
 425         /* check for a softlockup
 426          * This is done by making sure a high priority task is
 427          * being scheduled.  The task touches the watchdog to
 428          * indicate it is getting cpu time.  If it hasn't then
 429          * this is a good indication some task is hogging the cpu
 430          */
 431         duration = is_softlockup(touch_ts);
 432         if (unlikely(duration)) {
 433                 /*
 434                  * If a virtual machine is stopped by the host it can look to
 435                  * the watchdog like a soft lockup, check to see if the host
 436                  * stopped the vm before we issue the warning
 437                  */
 438                 if (kvm_check_and_clear_guest_paused())
 439                         return HRTIMER_RESTART;
 440
 441                 /* only warn once */
 442                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 443                         /*
 444                          * When multiple processes are causing softlockups the
 445                          * softlockup detector only warns on the first one
 446                          * because the code relies on a full quiet cycle to
 447                          * re-arm.  The second process prevents the quiet cycle
 448                          * and never gets reported.  Use task pointers to detect
 449                          * this.
 450                          */
 451                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 452                             current) {
 453                                 __this_cpu_write(soft_watchdog_warn, false);
 454                                 __touch_watchdog();
 455                         }
 456                         return HRTIMER_RESTART;
 457                 }
 458
 459                 if (softlockup_all_cpu_backtrace) {
 460                         /* Prevent multiple soft-lockup reports if one cpu is already
 461                          * engaged in dumping cpu back traces
 462                          */
 463                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 464                                 /* Someone else will report us. Let's give up */
 465                                 __this_cpu_write(soft_watchdog_warn, true);
 466                                 return HRTIMER_RESTART;
 467                         }
 468                 }
 469
 470                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 471                         smp_processor_id(), duration,
 472                         current->comm, task_pid_nr(current));
 473                 __this_cpu_write(softlockup_task_ptr_saved, current);
 474                 print_modules();
 475                 print_irqtrace_events(current);
 476                 if (regs)
 477                         show_regs(regs);
 478                 else
 479                         dump_stack();
 480
 481                 if (softlockup_all_cpu_backtrace) {
 482                         /* Avoid generating two back traces for current
 483                          * given that one is already made above
 484                          */
 485                         trigger_allbutself_cpu_backtrace();
 486
 487                         clear_bit(0, &soft_lockup_nmi_warn);
 488                         /* Barrier to sync with other cpus */
 489                         smp_mb__after_atomic();
 490                 }
 491
 492                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 493                 if (softlockup_panic)
 494                         panic("softlockup: hung tasks");
 495                 __this_cpu_write(soft_watchdog_warn, true);
 496         } else
 497                 __this_cpu_write(soft_watchdog_warn, false);
 498
 499         return HRTIMER_RESTART;
 500 }
 501
 502 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 503 {
 504         struct sched_param param = { .sched_priority = prio };
 505
 506         sched_setscheduler(current, policy, &param);
 507 }
 508
 509 static void watchdog_enable(unsigned int cpu)
 510 {
 511         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 512
 513         /* kick off the timer for the hardlockup detector */
 514         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 515         hrtimer->function = watchdog_timer_fn;
 516
 517         /* Enable the perf event */
 518         watchdog_nmi_enable(cpu);
 519
 520         /* done here because hrtimer_start can only pin to smp_processor_id() */
 521         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 522                       HRTIMER_MODE_REL_PINNED);
 523
 524         /* initialize timestamp */
 525         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 526         __touch_watchdog();
 527 }
 528
 529 static void watchdog_disable(unsigned int cpu)
 530 {
 531         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 532
 533         watchdog_set_prio(SCHED_NORMAL, 0);
 534         hrtimer_cancel(hrtimer);
 535         /* disable the perf event */
 536         watchdog_nmi_disable(cpu);
 537 }
 538
 539 static void watchdog_cleanup(unsigned int cpu, bool online)
 540 {
 541         watchdog_disable(cpu);
 542 }
 543
 544 static int watchdog_should_run(unsigned int cpu)
 545 {
 546         return __this_cpu_read(hrtimer_interrupts) !=
 547                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 548 }
 549
 550 /*
 551  * The watchdog thread function - touches the timestamp.
 552  *
 553  * It only runs once every sample_period seconds (4 seconds by
 554  * default) to reset the softlockup timestamp. If this gets delayed
 555  * for more than 2*watchdog_thresh seconds then the debug-printout
 556  * triggers in watchdog_timer_fn().
 557  */
 558 static void watchdog(unsigned int cpu)
 559 {
 560         __this_cpu_write(soft_lockup_hrtimer_cnt,
 561                          __this_cpu_read(hrtimer_interrupts));
 562         __touch_watchdog();
 563
 564         /*
 565          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 566          * failure path. Check for failures that can occur asynchronously -
 567          * for example, when CPUs are on-lined - and shut down the hardware
 568          * perf event on each CPU accordingly.
 569          *
 570          * The only non-obvious place this bit can be cleared is through
 571          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 572          * pr_info here would be too noisy as it would result in a message
 573          * every few seconds if the hardlockup was disabled but the softlockup
 574          * enabled.
 575          */
 576         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 577                 watchdog_nmi_disable(cpu);
 578 }
 579
 580 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 581 /*
 582  * People like the simple clean cpu node info on boot.
 583  * Reduce the watchdog noise by only printing messages
 584  * that are different from what cpu0 displayed.
 585  */
 586 static unsigned long cpu0_err;
 587
 588 static int watchdog_nmi_enable(unsigned int cpu)
 589 {
 590         struct perf_event_attr *wd_attr;
 591         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 592
 593         /* nothing to do if the hard lockup detector is disabled */
 594         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 595                 goto out;
 596
 597         /* is it already setup and enabled? */
 598         if (event && event->state > PERF_EVENT_STATE_OFF)
 599                 goto out;
 600
 601         /* it is setup but not enabled */
 602         if (event != NULL)
 603                 goto out_enable;
 604
 605         wd_attr = &wd_hw_attr;
 606         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 607
 608         /* Try to register using hardware perf events */
 609         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 610
 611         /* save cpu0 error for future comparision */
 612         if (cpu == 0 && IS_ERR(event))
 613                 cpu0_err = PTR_ERR(event);
 614
 615         if (!IS_ERR(event)) {
 616                 /* only print for cpu0 or different than cpu0 */
 617                 if (cpu == 0 || cpu0_err)
 618                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 619                 goto out_save;
 620         }
 621
 622         /*
 623          * Disable the hard lockup detector if _any_ CPU fails to set up
 624          * set up the hardware perf event. The watchdog() function checks
 625          * the NMI_WATCHDOG_ENABLED bit periodically.
 626          *
 627          * The barriers are for syncing up watchdog_enabled across all the
 628          * cpus, as clear_bit() does not use barriers.
 629          */
 630         smp_mb__before_atomic();
 631         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 632         smp_mb__after_atomic();
 633
 634         /* skip displaying the same error again */
 635         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 636                 return PTR_ERR(event);
 637
 638         /* vary the KERN level based on the returned errno */
 639         if (PTR_ERR(event) == -EOPNOTSUPP)
 640                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 641         else if (PTR_ERR(event) == -ENOENT)
 642                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 643                          cpu);
 644         else
 645                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 646                         cpu, PTR_ERR(event));
 647
 648         pr_info("Shutting down hard lockup detector on all cpus\n");
 649
 650         return PTR_ERR(event);
 651
 652         /* success path */
 653 out_save:
 654         per_cpu(watchdog_ev, cpu) = event;
 655 out_enable:
 656         perf_event_enable(per_cpu(watchdog_ev, cpu));
 657 out:
 658         return 0;
 659 }
 660
 661 static void watchdog_nmi_disable(unsigned int cpu)
 662 {
 663         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 664
 665         if (event) {
 666                 perf_event_disable(event);
 667                 per_cpu(watchdog_ev, cpu) = NULL;
 668
 669                 /* should be in cleanup, but blocks oprofile */
 670                 perf_event_release_kernel(event);
 671         }
 672         if (cpu == 0) {
 673                 /* watchdog_nmi_enable() expects this to be zero initially. */
 674                 cpu0_err = 0;
 675         }
 676 }
 677
 678 #else
 679 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 680 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 681 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 682
 683 static struct smp_hotplug_thread watchdog_threads = {
 684         .store                  = &softlockup_watchdog,
 685         .thread_should_run      = watchdog_should_run,
 686         .thread_fn              = watchdog,
 687         .thread_comm            = "watchdog/%u",
 688         .setup                  = watchdog_enable,
 689         .cleanup                = watchdog_cleanup,
 690         .park                   = watchdog_disable,
 691         .unpark                 = watchdog_enable,
 692 };
 693
 694 /*
 695  * park all watchdog threads that are specified in 'watchdog_cpumask'
 696  *
 697  * This function returns an error if kthread_park() of a watchdog thread
 698  * fails. In this situation, the watchdog threads of some CPUs can already
 699  * be parked and the watchdog threads of other CPUs can still be runnable.
 700  * Callers are expected to handle this special condition as appropriate in
 701  * their context.
 702  *
 703  * This function may only be called in a context that is protected against
 704  * races with CPU hotplug - for example, via get_online_cpus().
 705  */
 706 static int watchdog_park_threads(void)
 707 {
 708         int cpu, ret = 0;
 709
 710         for_each_watchdog_cpu(cpu) {
 711                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 712                 if (ret)
 713                         break;
 714         }
 715
 716         return ret;
 717 }
 718
 719 /*
 720  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 721  *
 722  * This function may only be called in a context that is protected against
 723  * races with CPU hotplug - for example, via get_online_cpus().
 724  */
 725 static void watchdog_unpark_threads(void)
 726 {
 727         int cpu;
 728
 729         for_each_watchdog_cpu(cpu)
 730                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 731 }
 732
 733 /*
 734  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 735  */
 736 int lockup_detector_suspend(void)
 737 {
 738         int ret = 0;
 739
 740         get_online_cpus();
 741         mutex_lock(&watchdog_proc_mutex);
 742         /*
 743          * Multiple suspend requests can be active in parallel (counted by
 744          * the 'watchdog_suspended' variable). If the watchdog threads are
 745          * running, the first caller takes care that they will be parked.
 746          * The state of 'watchdog_running' cannot change while a suspend
 747          * request is active (see related code in 'proc' handlers).
 748          */
 749         if (watchdog_running && !watchdog_suspended)
 750                 ret = watchdog_park_threads();
 751
 752         if (ret == 0)
 753                 watchdog_suspended++;
 754         else {
 755                 watchdog_disable_all_cpus();
 756                 pr_err("Failed to suspend lockup detectors, disabled\n");
 757                 watchdog_enabled = 0;
 758         }
 759
 760         mutex_unlock(&watchdog_proc_mutex);
 761
 762         return ret;
 763 }
 764
 765 /*
 766  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 767  */
 768 void lockup_detector_resume(void)
 769 {
 770         mutex_lock(&watchdog_proc_mutex);
 771
 772         watchdog_suspended--;
 773         /*
 774          * The watchdog threads are unparked if they were previously running
 775          * and if there is no more active suspend request.
 776          */
 777         if (watchdog_running && !watchdog_suspended)
 778                 watchdog_unpark_threads();
 779
 780         mutex_unlock(&watchdog_proc_mutex);
 781         put_online_cpus();
 782 }
 783
 784 static int update_watchdog_all_cpus(void)
 785 {
 786         int ret;
 787
 788         ret = watchdog_park_threads();
 789         if (ret)
 790                 return ret;
 791
 792         watchdog_unpark_threads();
 793
 794         return 0;
 795 }
 796
 797 static int watchdog_enable_all_cpus(void)
 798 {
 799         int err = 0;
 800
 801         if (!watchdog_running) {
 802                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 803                                                              &watchdog_cpumask);
 804                 if (err)
 805                         pr_err("Failed to create watchdog threads, disabled\n");
 806                 else
 807                         watchdog_running = 1;
 808         } else {
 809                 /*
 810                  * Enable/disable the lockup detectors or
 811                  * change the sample period 'on the fly'.
 812                  */
 813                 err = update_watchdog_all_cpus();
 814
 815                 if (err) {
 816                         watchdog_disable_all_cpus();
 817                         pr_err("Failed to update lockup detectors, disabled\n");
 818                 }
 819         }
 820
 821         if (err)
 822                 watchdog_enabled = 0;
 823
 824         return err;
 825 }
 826
 827 static void watchdog_disable_all_cpus(void)
 828 {
 829         if (watchdog_running) {
 830                 watchdog_running = 0;
 831                 smpboot_unregister_percpu_thread(&watchdog_threads);
 832         }
 833 }
 834
 835 #ifdef CONFIG_SYSCTL
 836
 837 /*
 838  * Update the run state of the lockup detectors.
 839  */
 840 static int proc_watchdog_update(void)
 841 {
 842         int err = 0;
 843
 844         /*
 845          * Watchdog threads won't be started if they are already active.
 846          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 847          * care of this. If those threads are already active, the sample
 848          * period will be updated and the lockup detectors will be enabled
 849          * or disabled 'on the fly'.
 850          */
 851         if (watchdog_enabled && watchdog_thresh)
 852                 err = watchdog_enable_all_cpus();
 853         else
 854                 watchdog_disable_all_cpus();
 855
 856         return err;
 857
 858 }
 859
 860 /*
 861  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 862  *
 863  * caller             | table->data points to | 'which' contains the flag(s)
 864  * -------------------|-----------------------|-----------------------------
 865  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 866  *                    |                       | with SOFT_WATCHDOG_ENABLED
 867  * -------------------|-----------------------|-----------------------------
 868  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 869  * -------------------|-----------------------|-----------------------------
 870  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 871  */
 872 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 873                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 874 {
 875         int err, old, new;
 876         int *watchdog_param = (int *)table->data;
 877
 878         get_online_cpus();
 879         mutex_lock(&watchdog_proc_mutex);
 880
 881         if (watchdog_suspended) {
 882                 /* no parameter changes allowed while watchdog is suspended */
 883                 err = -EAGAIN;
 884                 goto out;
 885         }
 886
 887         /*
 888          * If the parameter is being read return the state of the corresponding
 889          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 890          * run state of the lockup detectors.
 891          */
 892         if (!write) {
 893                 *watchdog_param = (watchdog_enabled & which) != 0;
 894                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 895         } else {
 896                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 897                 if (err)
 898                         goto out;
 899
 900                 /*
 901                  * There is a race window between fetching the current value
 902                  * from 'watchdog_enabled' and storing the new value. During
 903                  * this race window, watchdog_nmi_enable() can sneak in and
 904                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 905                  * The 'cmpxchg' detects this race and the loop retries.
 906                  */
 907                 do {
 908                         old = watchdog_enabled;
 909                         /*
 910                          * If the parameter value is not zero set the
 911                          * corresponding bit(s), else clear it(them).
 912                          */
 913                         if (*watchdog_param)
 914                                 new = old | which;
 915                         else
 916                                 new = old & ~which;
 917                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 918
 919                 /*
 920                  * Update the run state of the lockup detectors. There is _no_
 921                  * need to check the value returned by proc_watchdog_update()
 922                  * and to restore the previous value of 'watchdog_enabled' as
 923                  * both lockup detectors are disabled if proc_watchdog_update()
 924                  * returns an error.
 925                  */
 926                 if (old == new)
 927                         goto out;
 928
 929                 err = proc_watchdog_update();
 930         }
 931 out:
 932         mutex_unlock(&watchdog_proc_mutex);
 933         put_online_cpus();
 934         return err;
 935 }
 936
 937 /*
 938  * /proc/sys/kernel/watchdog
 939  */
 940 int proc_watchdog(struct ctl_table *table, int write,
 941                   void __user *buffer, size_t *lenp, loff_t *ppos)
 942 {
 943         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 944                                     table, write, buffer, lenp, ppos);
 945 }
 946
 947 /*
 948  * /proc/sys/kernel/nmi_watchdog
 949  */
 950 int proc_nmi_watchdog(struct ctl_table *table, int write,
 951                       void __user *buffer, size_t *lenp, loff_t *ppos)
 952 {
 953         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 954                                     table, write, buffer, lenp, ppos);
 955 }
 956
 957 /*
 958  * /proc/sys/kernel/soft_watchdog
 959  */
 960 int proc_soft_watchdog(struct ctl_table *table, int write,
 961                         void __user *buffer, size_t *lenp, loff_t *ppos)
 962 {
 963         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 964                                     table, write, buffer, lenp, ppos);
 965 }
 966
 967 /*
 968  * /proc/sys/kernel/watchdog_thresh
 969  */
 970 int proc_watchdog_thresh(struct ctl_table *table, int write,
 971                          void __user *buffer, size_t *lenp, loff_t *ppos)
 972 {
 973         int err, old, new;
 974
 975         get_online_cpus();
 976         mutex_lock(&watchdog_proc_mutex);
 977
 978         if (watchdog_suspended) {
 979                 /* no parameter changes allowed while watchdog is suspended */
 980                 err = -EAGAIN;
 981                 goto out;
 982         }
 983
 984         old = ACCESS_ONCE(watchdog_thresh);
 985         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 986
 987         if (err || !write)
 988                 goto out;
 989
 990         /*
 991          * Update the sample period. Restore on failure.
 992          */
 993         new = ACCESS_ONCE(watchdog_thresh);
 994         if (old == new)
 995                 goto out;
 996
 997         set_sample_period();
 998         err = proc_watchdog_update();
 999         if (err) {
1000                 watchdog_thresh = old;
1001                 set_sample_period();
1002         }
1003 out:
1004         mutex_unlock(&watchdog_proc_mutex);
1005         put_online_cpus();
1006         return err;
1007 }
1008
1009 /*
1010  * The cpumask is the mask of possible cpus that the watchdog can run
1011  * on, not the mask of cpus it is actually running on.  This allows the
1012  * user to specify a mask that will include cpus that have not yet
1013  * been brought online, if desired.
1014  */
1015 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1016                           void __user *buffer, size_t *lenp, loff_t *ppos)
1017 {
1018         int err;
1019
1020         get_online_cpus();
1021         mutex_lock(&watchdog_proc_mutex);
1022
1023         if (watchdog_suspended) {
1024                 /* no parameter changes allowed while watchdog is suspended */
1025                 err = -EAGAIN;
1026                 goto out;
1027         }
1028
1029         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1030         if (!err && write) {
1031                 /* Remove impossible cpus to keep sysctl output cleaner. */
1032                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1033                             cpu_possible_mask);
1034
1035                 if (watchdog_running) {
1036                         /*
1037                          * Failure would be due to being unable to allocate
1038                          * a temporary cpumask, so we are likely not in a
1039                          * position to do much else to make things better.
1040                          */
1041                         if (smpboot_update_cpumask_percpu_thread(
1042                                     &watchdog_threads, &watchdog_cpumask) != 0)
1043                                 pr_err("cpumask update failed\n");
1044                 }
1045         }
1046 out:
1047         mutex_unlock(&watchdog_proc_mutex);
1048         put_online_cpus();
1049         return err;
1050 }
1051
1052 #endif /* CONFIG_SYSCTL */
1053
1054 void __init lockup_detector_init(void)
1055 {
1056         set_sample_period();
1057
1058 #ifdef CONFIG_NO_HZ_FULL
1059         if (tick_nohz_full_enabled()) {
1060                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1061                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1062         } else
1063                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1064 #else
1065         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1066 #endif
1067
1068         if (watchdog_enabled)
1069                 watchdog_enable_all_cpus();
1070 }