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 struct perf_event_attr wd_hw_attr = {
 303         .type           = PERF_TYPE_HARDWARE,
 304         .config         = PERF_COUNT_HW_CPU_CYCLES,
 305         .size           = sizeof(struct perf_event_attr),
 306         .pinned         = 1,
 307         .disabled       = 1,
 308 };
 309
 310 /* Callback function for perf event subsystem */
 311 static void watchdog_overflow_callback(struct perf_event *event,
 312                  struct perf_sample_data *data,
 313                  struct pt_regs *regs)
 314 {
 315         /* Ensure the watchdog never gets throttled */
 316         event->hw.interrupts = 0;
 317
 318         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 319                 __this_cpu_write(watchdog_nmi_touch, false);
 320                 return;
 321         }
 322
 323         /* check for a hardlockup
 324          * This is done by making sure our timer interrupt
 325          * is incrementing.  The timer interrupt should have
 326          * fired multiple times before we overflow'd.  If it hasn't
 327          * then this is a good indication the cpu is stuck
 328          */
 329         if (is_hardlockup()) {
 330                 int this_cpu = smp_processor_id();
 331                 struct pt_regs *regs = get_irq_regs();
 332
 333                 /* only print hardlockups once */
 334                 if (__this_cpu_read(hard_watchdog_warn) == true)
 335                         return;
 336
 337                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 338                 print_modules();
 339                 print_irqtrace_events(current);
 340                 if (regs)
 341                         show_regs(regs);
 342                 else
 343                         dump_stack();
 344
 345                 /*
 346                  * Perform all-CPU dump only once to avoid multiple hardlockups
 347                  * generating interleaving traces
 348                  */
 349                 if (sysctl_hardlockup_all_cpu_backtrace &&
 350                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 351                         trigger_allbutself_cpu_backtrace();
 352
 353                 if (hardlockup_panic)
 354                         panic("Hard LOCKUP");
 355
 356                 __this_cpu_write(hard_watchdog_warn, true);
 357                 return;
 358         }
 359
 360         __this_cpu_write(hard_watchdog_warn, false);
 361         return;
 362 }
 363 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 364
 365 static void watchdog_interrupt_count(void)
 366 {
 367         __this_cpu_inc(hrtimer_interrupts);
 368 }
 369
 370 static int watchdog_nmi_enable(unsigned int cpu);
 371 static void watchdog_nmi_disable(unsigned int cpu);
 372
 373 static int watchdog_enable_all_cpus(void);
 374 static void watchdog_disable_all_cpus(void);
 375
 376 /* watchdog kicker functions */
 377 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 378 {
 379         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 380         struct pt_regs *regs = get_irq_regs();
 381         int duration;
 382         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 383
 384         /* kick the hardlockup detector */
 385         watchdog_interrupt_count();
 386
 387         /* kick the softlockup detector */
 388         wake_up_process(__this_cpu_read(softlockup_watchdog));
 389
 390         /* .. and repeat */
 391         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 392
 393         if (touch_ts == 0) {
 394                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 395                         /*
 396                          * If the time stamp was touched atomically
 397                          * make sure the scheduler tick is up to date.
 398                          */
 399                         __this_cpu_write(softlockup_touch_sync, false);
 400                         sched_clock_tick();
 401                 }
 402
 403                 /* Clear the guest paused flag on watchdog reset */
 404                 kvm_check_and_clear_guest_paused();
 405                 __touch_watchdog();
 406                 return HRTIMER_RESTART;
 407         }
 408
 409         /* check for a softlockup
 410          * This is done by making sure a high priority task is
 411          * being scheduled.  The task touches the watchdog to
 412          * indicate it is getting cpu time.  If it hasn't then
 413          * this is a good indication some task is hogging the cpu
 414          */
 415         duration = is_softlockup(touch_ts);
 416         if (unlikely(duration)) {
 417                 /*
 418                  * If a virtual machine is stopped by the host it can look to
 419                  * the watchdog like a soft lockup, check to see if the host
 420                  * stopped the vm before we issue the warning
 421                  */
 422                 if (kvm_check_and_clear_guest_paused())
 423                         return HRTIMER_RESTART;
 424
 425                 /* only warn once */
 426                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 427                         /*
 428                          * When multiple processes are causing softlockups the
 429                          * softlockup detector only warns on the first one
 430                          * because the code relies on a full quiet cycle to
 431                          * re-arm.  The second process prevents the quiet cycle
 432                          * and never gets reported.  Use task pointers to detect
 433                          * this.
 434                          */
 435                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 436                             current) {
 437                                 __this_cpu_write(soft_watchdog_warn, false);
 438                                 __touch_watchdog();
 439                         }
 440                         return HRTIMER_RESTART;
 441                 }
 442
 443                 if (softlockup_all_cpu_backtrace) {
 444                         /* Prevent multiple soft-lockup reports if one cpu is already
 445                          * engaged in dumping cpu back traces
 446                          */
 447                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 448                                 /* Someone else will report us. Let's give up */
 449                                 __this_cpu_write(soft_watchdog_warn, true);
 450                                 return HRTIMER_RESTART;
 451                         }
 452                 }
 453
 454                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 455                         smp_processor_id(), duration,
 456                         current->comm, task_pid_nr(current));
 457                 __this_cpu_write(softlockup_task_ptr_saved, current);
 458                 print_modules();
 459                 print_irqtrace_events(current);
 460                 if (regs)
 461                         show_regs(regs);
 462                 else
 463                         dump_stack();
 464
 465                 if (softlockup_all_cpu_backtrace) {
 466                         /* Avoid generating two back traces for current
 467                          * given that one is already made above
 468                          */
 469                         trigger_allbutself_cpu_backtrace();
 470
 471                         clear_bit(0, &soft_lockup_nmi_warn);
 472                         /* Barrier to sync with other cpus */
 473                         smp_mb__after_atomic();
 474                 }
 475
 476                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 477                 if (softlockup_panic)
 478                         panic("softlockup: hung tasks");
 479                 __this_cpu_write(soft_watchdog_warn, true);
 480         } else
 481                 __this_cpu_write(soft_watchdog_warn, false);
 482
 483         return HRTIMER_RESTART;
 484 }
 485
 486 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 487 {
 488         struct sched_param param = { .sched_priority = prio };
 489
 490         sched_setscheduler(current, policy, &param);
 491 }
 492
 493 static void watchdog_enable(unsigned int cpu)
 494 {
 495         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 496
 497         /* kick off the timer for the hardlockup detector */
 498         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 499         hrtimer->function = watchdog_timer_fn;
 500
 501         /* Enable the perf event */
 502         watchdog_nmi_enable(cpu);
 503
 504         /* done here because hrtimer_start can only pin to smp_processor_id() */
 505         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 506                       HRTIMER_MODE_REL_PINNED);
 507
 508         /* initialize timestamp */
 509         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 510         __touch_watchdog();
 511 }
 512
 513 static void watchdog_disable(unsigned int cpu)
 514 {
 515         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 516
 517         watchdog_set_prio(SCHED_NORMAL, 0);
 518         hrtimer_cancel(hrtimer);
 519         /* disable the perf event */
 520         watchdog_nmi_disable(cpu);
 521 }
 522
 523 static void watchdog_cleanup(unsigned int cpu, bool online)
 524 {
 525         watchdog_disable(cpu);
 526 }
 527
 528 static int watchdog_should_run(unsigned int cpu)
 529 {
 530         return __this_cpu_read(hrtimer_interrupts) !=
 531                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 532 }
 533
 534 /*
 535  * The watchdog thread function - touches the timestamp.
 536  *
 537  * It only runs once every sample_period seconds (4 seconds by
 538  * default) to reset the softlockup timestamp. If this gets delayed
 539  * for more than 2*watchdog_thresh seconds then the debug-printout
 540  * triggers in watchdog_timer_fn().
 541  */
 542 static void watchdog(unsigned int cpu)
 543 {
 544         __this_cpu_write(soft_lockup_hrtimer_cnt,
 545                          __this_cpu_read(hrtimer_interrupts));
 546         __touch_watchdog();
 547
 548         /*
 549          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 550          * failure path. Check for failures that can occur asynchronously -
 551          * for example, when CPUs are on-lined - and shut down the hardware
 552          * perf event on each CPU accordingly.
 553          *
 554          * The only non-obvious place this bit can be cleared is through
 555          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 556          * pr_info here would be too noisy as it would result in a message
 557          * every few seconds if the hardlockup was disabled but the softlockup
 558          * enabled.
 559          */
 560         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 561                 watchdog_nmi_disable(cpu);
 562 }
 563
 564 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 565 /*
 566  * People like the simple clean cpu node info on boot.
 567  * Reduce the watchdog noise by only printing messages
 568  * that are different from what cpu0 displayed.
 569  */
 570 static unsigned long cpu0_err;
 571
 572 static int watchdog_nmi_enable(unsigned int cpu)
 573 {
 574         struct perf_event_attr *wd_attr;
 575         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 576
 577         /* nothing to do if the hard lockup detector is disabled */
 578         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 579                 goto out;
 580
 581         /* is it already setup and enabled? */
 582         if (event && event->state > PERF_EVENT_STATE_OFF)
 583                 goto out;
 584
 585         /* it is setup but not enabled */
 586         if (event != NULL)
 587                 goto out_enable;
 588
 589         wd_attr = &wd_hw_attr;
 590         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 591
 592         /* Try to register using hardware perf events */
 593         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 594
 595         /* save cpu0 error for future comparision */
 596         if (cpu == 0 && IS_ERR(event))
 597                 cpu0_err = PTR_ERR(event);
 598
 599         if (!IS_ERR(event)) {
 600                 /* only print for cpu0 or different than cpu0 */
 601                 if (cpu == 0 || cpu0_err)
 602                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 603                 goto out_save;
 604         }
 605
 606         /*
 607          * Disable the hard lockup detector if _any_ CPU fails to set up
 608          * set up the hardware perf event. The watchdog() function checks
 609          * the NMI_WATCHDOG_ENABLED bit periodically.
 610          *
 611          * The barriers are for syncing up watchdog_enabled across all the
 612          * cpus, as clear_bit() does not use barriers.
 613          */
 614         smp_mb__before_atomic();
 615         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 616         smp_mb__after_atomic();
 617
 618         /* skip displaying the same error again */
 619         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 620                 return PTR_ERR(event);
 621
 622         /* vary the KERN level based on the returned errno */
 623         if (PTR_ERR(event) == -EOPNOTSUPP)
 624                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 625         else if (PTR_ERR(event) == -ENOENT)
 626                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 627                          cpu);
 628         else
 629                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 630                         cpu, PTR_ERR(event));
 631
 632         pr_info("Shutting down hard lockup detector on all cpus\n");
 633
 634         return PTR_ERR(event);
 635
 636         /* success path */
 637 out_save:
 638         per_cpu(watchdog_ev, cpu) = event;
 639 out_enable:
 640         perf_event_enable(per_cpu(watchdog_ev, cpu));
 641 out:
 642         return 0;
 643 }
 644
 645 static void watchdog_nmi_disable(unsigned int cpu)
 646 {
 647         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 648
 649         if (event) {
 650                 perf_event_disable(event);
 651                 per_cpu(watchdog_ev, cpu) = NULL;
 652
 653                 /* should be in cleanup, but blocks oprofile */
 654                 perf_event_release_kernel(event);
 655         }
 656         if (cpu == 0) {
 657                 /* watchdog_nmi_enable() expects this to be zero initially. */
 658                 cpu0_err = 0;
 659         }
 660 }
 661
 662 #else
 663 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 664 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 665 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 666
 667 static struct smp_hotplug_thread watchdog_threads = {
 668         .store                  = &softlockup_watchdog,
 669         .thread_should_run      = watchdog_should_run,
 670         .thread_fn              = watchdog,
 671         .thread_comm            = "watchdog/%u",
 672         .setup                  = watchdog_enable,
 673         .cleanup                = watchdog_cleanup,
 674         .park                   = watchdog_disable,
 675         .unpark                 = watchdog_enable,
 676 };
 677
 678 /*
 679  * park all watchdog threads that are specified in 'watchdog_cpumask'
 680  *
 681  * This function returns an error if kthread_park() of a watchdog thread
 682  * fails. In this situation, the watchdog threads of some CPUs can already
 683  * be parked and the watchdog threads of other CPUs can still be runnable.
 684  * Callers are expected to handle this special condition as appropriate in
 685  * their context.
 686  *
 687  * This function may only be called in a context that is protected against
 688  * races with CPU hotplug - for example, via get_online_cpus().
 689  */
 690 static int watchdog_park_threads(void)
 691 {
 692         int cpu, ret = 0;
 693
 694         for_each_watchdog_cpu(cpu) {
 695                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 696                 if (ret)
 697                         break;
 698         }
 699
 700         return ret;
 701 }
 702
 703 /*
 704  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 705  *
 706  * This function may only be called in a context that is protected against
 707  * races with CPU hotplug - for example, via get_online_cpus().
 708  */
 709 static void watchdog_unpark_threads(void)
 710 {
 711         int cpu;
 712
 713         for_each_watchdog_cpu(cpu)
 714                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 715 }
 716
 717 /*
 718  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 719  */
 720 int lockup_detector_suspend(void)
 721 {
 722         int ret = 0;
 723
 724         get_online_cpus();
 725         mutex_lock(&watchdog_proc_mutex);
 726         /*
 727          * Multiple suspend requests can be active in parallel (counted by
 728          * the 'watchdog_suspended' variable). If the watchdog threads are
 729          * running, the first caller takes care that they will be parked.
 730          * The state of 'watchdog_running' cannot change while a suspend
 731          * request is active (see related code in 'proc' handlers).
 732          */
 733         if (watchdog_running && !watchdog_suspended)
 734                 ret = watchdog_park_threads();
 735
 736         if (ret == 0)
 737                 watchdog_suspended++;
 738         else {
 739                 watchdog_disable_all_cpus();
 740                 pr_err("Failed to suspend lockup detectors, disabled\n");
 741                 watchdog_enabled = 0;
 742         }
 743
 744         mutex_unlock(&watchdog_proc_mutex);
 745
 746         return ret;
 747 }
 748
 749 /*
 750  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 751  */
 752 void lockup_detector_resume(void)
 753 {
 754         mutex_lock(&watchdog_proc_mutex);
 755
 756         watchdog_suspended--;
 757         /*
 758          * The watchdog threads are unparked if they were previously running
 759          * and if there is no more active suspend request.
 760          */
 761         if (watchdog_running && !watchdog_suspended)
 762                 watchdog_unpark_threads();
 763
 764         mutex_unlock(&watchdog_proc_mutex);
 765         put_online_cpus();
 766 }
 767
 768 static int update_watchdog_all_cpus(void)
 769 {
 770         int ret;
 771
 772         ret = watchdog_park_threads();
 773         if (ret)
 774                 return ret;
 775
 776         watchdog_unpark_threads();
 777
 778         return 0;
 779 }
 780
 781 static int watchdog_enable_all_cpus(void)
 782 {
 783         int err = 0;
 784
 785         if (!watchdog_running) {
 786                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 787                                                              &watchdog_cpumask);
 788                 if (err)
 789                         pr_err("Failed to create watchdog threads, disabled\n");
 790                 else
 791                         watchdog_running = 1;
 792         } else {
 793                 /*
 794                  * Enable/disable the lockup detectors or
 795                  * change the sample period 'on the fly'.
 796                  */
 797                 err = update_watchdog_all_cpus();
 798
 799                 if (err) {
 800                         watchdog_disable_all_cpus();
 801                         pr_err("Failed to update lockup detectors, disabled\n");
 802                 }
 803         }
 804
 805         if (err)
 806                 watchdog_enabled = 0;
 807
 808         return err;
 809 }
 810
 811 static void watchdog_disable_all_cpus(void)
 812 {
 813         if (watchdog_running) {
 814                 watchdog_running = 0;
 815                 smpboot_unregister_percpu_thread(&watchdog_threads);
 816         }
 817 }
 818
 819 #ifdef CONFIG_SYSCTL
 820
 821 /*
 822  * Update the run state of the lockup detectors.
 823  */
 824 static int proc_watchdog_update(void)
 825 {
 826         int err = 0;
 827
 828         /*
 829          * Watchdog threads won't be started if they are already active.
 830          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 831          * care of this. If those threads are already active, the sample
 832          * period will be updated and the lockup detectors will be enabled
 833          * or disabled 'on the fly'.
 834          */
 835         if (watchdog_enabled && watchdog_thresh)
 836                 err = watchdog_enable_all_cpus();
 837         else
 838                 watchdog_disable_all_cpus();
 839
 840         return err;
 841
 842 }
 843
 844 /*
 845  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 846  *
 847  * caller             | table->data points to | 'which' contains the flag(s)
 848  * -------------------|-----------------------|-----------------------------
 849  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 850  *                    |                       | with SOFT_WATCHDOG_ENABLED
 851  * -------------------|-----------------------|-----------------------------
 852  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 853  * -------------------|-----------------------|-----------------------------
 854  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 855  */
 856 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 857                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 858 {
 859         int err, old, new;
 860         int *watchdog_param = (int *)table->data;
 861
 862         get_online_cpus();
 863         mutex_lock(&watchdog_proc_mutex);
 864
 865         if (watchdog_suspended) {
 866                 /* no parameter changes allowed while watchdog is suspended */
 867                 err = -EAGAIN;
 868                 goto out;
 869         }
 870
 871         /*
 872          * If the parameter is being read return the state of the corresponding
 873          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 874          * run state of the lockup detectors.
 875          */
 876         if (!write) {
 877                 *watchdog_param = (watchdog_enabled & which) != 0;
 878                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 879         } else {
 880                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 881                 if (err)
 882                         goto out;
 883
 884                 /*
 885                  * There is a race window between fetching the current value
 886                  * from 'watchdog_enabled' and storing the new value. During
 887                  * this race window, watchdog_nmi_enable() can sneak in and
 888                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 889                  * The 'cmpxchg' detects this race and the loop retries.
 890                  */
 891                 do {
 892                         old = watchdog_enabled;
 893                         /*
 894                          * If the parameter value is not zero set the
 895                          * corresponding bit(s), else clear it(them).
 896                          */
 897                         if (*watchdog_param)
 898                                 new = old | which;
 899                         else
 900                                 new = old & ~which;
 901                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 902
 903                 /*
 904                  * Update the run state of the lockup detectors. There is _no_
 905                  * need to check the value returned by proc_watchdog_update()
 906                  * and to restore the previous value of 'watchdog_enabled' as
 907                  * both lockup detectors are disabled if proc_watchdog_update()
 908                  * returns an error.
 909                  */
 910                 err = proc_watchdog_update();
 911         }
 912 out:
 913         mutex_unlock(&watchdog_proc_mutex);
 914         put_online_cpus();
 915         return err;
 916 }
 917
 918 /*
 919  * /proc/sys/kernel/watchdog
 920  */
 921 int proc_watchdog(struct ctl_table *table, int write,
 922                   void __user *buffer, size_t *lenp, loff_t *ppos)
 923 {
 924         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 925                                     table, write, buffer, lenp, ppos);
 926 }
 927
 928 /*
 929  * /proc/sys/kernel/nmi_watchdog
 930  */
 931 int proc_nmi_watchdog(struct ctl_table *table, int write,
 932                       void __user *buffer, size_t *lenp, loff_t *ppos)
 933 {
 934         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 935                                     table, write, buffer, lenp, ppos);
 936 }
 937
 938 /*
 939  * /proc/sys/kernel/soft_watchdog
 940  */
 941 int proc_soft_watchdog(struct ctl_table *table, int write,
 942                         void __user *buffer, size_t *lenp, loff_t *ppos)
 943 {
 944         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 945                                     table, write, buffer, lenp, ppos);
 946 }
 947
 948 /*
 949  * /proc/sys/kernel/watchdog_thresh
 950  */
 951 int proc_watchdog_thresh(struct ctl_table *table, int write,
 952                          void __user *buffer, size_t *lenp, loff_t *ppos)
 953 {
 954         int err, old;
 955
 956         get_online_cpus();
 957         mutex_lock(&watchdog_proc_mutex);
 958
 959         if (watchdog_suspended) {
 960                 /* no parameter changes allowed while watchdog is suspended */
 961                 err = -EAGAIN;
 962                 goto out;
 963         }
 964
 965         old = ACCESS_ONCE(watchdog_thresh);
 966         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 967
 968         if (err || !write)
 969                 goto out;
 970
 971         /*
 972          * Update the sample period. Restore on failure.
 973          */
 974         set_sample_period();
 975         err = proc_watchdog_update();
 976         if (err) {
 977                 watchdog_thresh = old;
 978                 set_sample_period();
 979         }
 980 out:
 981         mutex_unlock(&watchdog_proc_mutex);
 982         put_online_cpus();
 983         return err;
 984 }
 985
 986 /*
 987  * The cpumask is the mask of possible cpus that the watchdog can run
 988  * on, not the mask of cpus it is actually running on.  This allows the
 989  * user to specify a mask that will include cpus that have not yet
 990  * been brought online, if desired.
 991  */
 992 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 993                           void __user *buffer, size_t *lenp, loff_t *ppos)
 994 {
 995         int err;
 996
 997         get_online_cpus();
 998         mutex_lock(&watchdog_proc_mutex);
 999
1000         if (watchdog_suspended) {
1001                 /* no parameter changes allowed while watchdog is suspended */
1002                 err = -EAGAIN;
1003                 goto out;
1004         }
1005
1006         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1007         if (!err && write) {
1008                 /* Remove impossible cpus to keep sysctl output cleaner. */
1009                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1010                             cpu_possible_mask);
1011
1012                 if (watchdog_running) {
1013                         /*
1014                          * Failure would be due to being unable to allocate
1015                          * a temporary cpumask, so we are likely not in a
1016                          * position to do much else to make things better.
1017                          */
1018                         if (smpboot_update_cpumask_percpu_thread(
1019                                     &watchdog_threads, &watchdog_cpumask) != 0)
1020                                 pr_err("cpumask update failed\n");
1021                 }
1022         }
1023 out:
1024         mutex_unlock(&watchdog_proc_mutex);
1025         put_online_cpus();
1026         return err;
1027 }
1028
1029 #endif /* CONFIG_SYSCTL */
1030
1031 void __init lockup_detector_init(void)
1032 {
1033         set_sample_period();
1034
1035 #ifdef CONFIG_NO_HZ_FULL
1036         if (tick_nohz_full_enabled()) {
1037                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1038                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1039         } else
1040                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1041 #else
1042         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1043 #endif
1044
1045         if (watchdog_enabled)
1046                 watchdog_enable_all_cpus();
1047 }