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