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 DEFINE_RAW_SPINLOCK(watchdog_output_lock);
 319
 320 static struct perf_event_attr wd_hw_attr = {
 321         .type           = PERF_TYPE_HARDWARE,
 322         .config         = PERF_COUNT_HW_CPU_CYCLES,
 323         .size           = sizeof(struct perf_event_attr),
 324         .pinned         = 1,
 325         .disabled       = 1,
 326 };
 327
 328 /* Callback function for perf event subsystem */
 329 static void watchdog_overflow_callback(struct perf_event *event,
 330                  struct perf_sample_data *data,
 331                  struct pt_regs *regs)
 332 {
 333         /* Ensure the watchdog never gets throttled */
 334         event->hw.interrupts = 0;
 335
 336         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 337                 __this_cpu_write(watchdog_nmi_touch, false);
 338                 return;
 339         }
 340
 341         /* check for a hardlockup
 342          * This is done by making sure our timer interrupt
 343          * is incrementing.  The timer interrupt should have
 344          * fired multiple times before we overflow'd.  If it hasn't
 345          * then this is a good indication the cpu is stuck
 346          */
 347         if (is_hardlockup()) {
 348                 int this_cpu = smp_processor_id();
 349                 struct pt_regs *regs = get_irq_regs();
 350
 351                 /* only print hardlockups once */
 352                 if (__this_cpu_read(hard_watchdog_warn) == true)
 353                         return;
 354                 /*
 355                  * If early-printk is enabled then make sure we do not
 356                  * lock up in printk() and kill console logging:
 357                  */
 358                 printk_kill();
 359
 360                 raw_spin_lock(&watchdog_output_lock);
 361
 362                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 363                 print_modules();
 364                 print_irqtrace_events(current);
 365                 if (regs)
 366                         show_regs(regs);
 367                 else
 368                         dump_stack();
 369
 370                 /*
 371                  * Perform all-CPU dump only once to avoid multiple hardlockups
 372                  * generating interleaving traces
 373                  */
 374                 if (sysctl_hardlockup_all_cpu_backtrace &&
 375                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 376                         trigger_allbutself_cpu_backtrace();
 377
 378                 raw_spin_unlock(&watchdog_output_lock);
 379                 if (hardlockup_panic)
 380                         nmi_panic(regs, "Hard LOCKUP");
 381
 382                 __this_cpu_write(hard_watchdog_warn, true);
 383                 return;
 384         }
 385
 386         __this_cpu_write(hard_watchdog_warn, false);
 387         return;
 388 }
 389 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 390
 391 static void watchdog_interrupt_count(void)
 392 {
 393         __this_cpu_inc(hrtimer_interrupts);
 394 }
 395
 396 static int watchdog_nmi_enable(unsigned int cpu);
 397 static void watchdog_nmi_disable(unsigned int cpu);
 398
 399 static int watchdog_enable_all_cpus(void);
 400 static void watchdog_disable_all_cpus(void);
 401
 402 /* watchdog kicker functions */
 403 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 404 {
 405         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 406         struct pt_regs *regs = get_irq_regs();
 407         int duration;
 408         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 409
 410         /* kick the hardlockup detector */
 411         watchdog_interrupt_count();
 412
 413         /* kick the softlockup detector */
 414         wake_up_process(__this_cpu_read(softlockup_watchdog));
 415
 416         /* .. and repeat */
 417         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 418
 419         if (touch_ts == 0) {
 420                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 421                         /*
 422                          * If the time stamp was touched atomically
 423                          * make sure the scheduler tick is up to date.
 424                          */
 425                         __this_cpu_write(softlockup_touch_sync, false);
 426                         sched_clock_tick();
 427                 }
 428
 429                 /* Clear the guest paused flag on watchdog reset */
 430                 kvm_check_and_clear_guest_paused();
 431                 __touch_watchdog();
 432                 return HRTIMER_RESTART;
 433         }
 434
 435         /* check for a softlockup
 436          * This is done by making sure a high priority task is
 437          * being scheduled.  The task touches the watchdog to
 438          * indicate it is getting cpu time.  If it hasn't then
 439          * this is a good indication some task is hogging the cpu
 440          */
 441         duration = is_softlockup(touch_ts);
 442         if (unlikely(duration)) {
 443                 /*
 444                  * If a virtual machine is stopped by the host it can look to
 445                  * the watchdog like a soft lockup, check to see if the host
 446                  * stopped the vm before we issue the warning
 447                  */
 448                 if (kvm_check_and_clear_guest_paused())
 449                         return HRTIMER_RESTART;
 450
 451                 /* only warn once */
 452                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 453                         /*
 454                          * When multiple processes are causing softlockups the
 455                          * softlockup detector only warns on the first one
 456                          * because the code relies on a full quiet cycle to
 457                          * re-arm.  The second process prevents the quiet cycle
 458                          * and never gets reported.  Use task pointers to detect
 459                          * this.
 460                          */
 461                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 462                             current) {
 463                                 __this_cpu_write(soft_watchdog_warn, false);
 464                                 __touch_watchdog();
 465                         }
 466                         return HRTIMER_RESTART;
 467                 }
 468
 469                 if (softlockup_all_cpu_backtrace) {
 470                         /* Prevent multiple soft-lockup reports if one cpu is already
 471                          * engaged in dumping cpu back traces
 472                          */
 473                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 474                                 /* Someone else will report us. Let's give up */
 475                                 __this_cpu_write(soft_watchdog_warn, true);
 476                                 return HRTIMER_RESTART;
 477                         }
 478                 }
 479
 480                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 481                         smp_processor_id(), duration,
 482                         current->comm, task_pid_nr(current));
 483                 __this_cpu_write(softlockup_task_ptr_saved, current);
 484                 print_modules();
 485                 print_irqtrace_events(current);
 486                 if (regs)
 487                         show_regs(regs);
 488                 else
 489                         dump_stack();
 490
 491                 if (softlockup_all_cpu_backtrace) {
 492                         /* Avoid generating two back traces for current
 493                          * given that one is already made above
 494                          */
 495                         trigger_allbutself_cpu_backtrace();
 496
 497                         clear_bit(0, &soft_lockup_nmi_warn);
 498                         /* Barrier to sync with other cpus */
 499                         smp_mb__after_atomic();
 500                 }
 501
 502                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 503                 if (softlockup_panic)
 504                         panic("softlockup: hung tasks");
 505                 __this_cpu_write(soft_watchdog_warn, true);
 506         } else
 507                 __this_cpu_write(soft_watchdog_warn, false);
 508
 509         return HRTIMER_RESTART;
 510 }
 511
 512 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 513 {
 514         struct sched_param param = { .sched_priority = prio };
 515
 516         sched_setscheduler(current, policy, &param);
 517 }
 518
 519 static void watchdog_enable(unsigned int cpu)
 520 {
 521         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 522
 523         /* kick off the timer for the hardlockup detector */
 524         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 525         hrtimer->function = watchdog_timer_fn;
 526         hrtimer->irqsafe = 1;
 527
 528         /* Enable the perf event */
 529         watchdog_nmi_enable(cpu);
 530
 531         /* done here because hrtimer_start can only pin to smp_processor_id() */
 532         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 533                       HRTIMER_MODE_REL_PINNED);
 534
 535         /* initialize timestamp */
 536         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 537         __touch_watchdog();
 538 }
 539
 540 static void watchdog_disable(unsigned int cpu)
 541 {
 542         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 543
 544         watchdog_set_prio(SCHED_NORMAL, 0);
 545         hrtimer_cancel(hrtimer);
 546         /* disable the perf event */
 547         watchdog_nmi_disable(cpu);
 548 }
 549
 550 static void watchdog_cleanup(unsigned int cpu, bool online)
 551 {
 552         watchdog_disable(cpu);
 553 }
 554
 555 static int watchdog_should_run(unsigned int cpu)
 556 {
 557         return __this_cpu_read(hrtimer_interrupts) !=
 558                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 559 }
 560
 561 /*
 562  * The watchdog thread function - touches the timestamp.
 563  *
 564  * It only runs once every sample_period seconds (4 seconds by
 565  * default) to reset the softlockup timestamp. If this gets delayed
 566  * for more than 2*watchdog_thresh seconds then the debug-printout
 567  * triggers in watchdog_timer_fn().
 568  */
 569 static void watchdog(unsigned int cpu)
 570 {
 571         __this_cpu_write(soft_lockup_hrtimer_cnt,
 572                          __this_cpu_read(hrtimer_interrupts));
 573         __touch_watchdog();
 574
 575         /*
 576          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 577          * failure path. Check for failures that can occur asynchronously -
 578          * for example, when CPUs are on-lined - and shut down the hardware
 579          * perf event on each CPU accordingly.
 580          *
 581          * The only non-obvious place this bit can be cleared is through
 582          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 583          * pr_info here would be too noisy as it would result in a message
 584          * every few seconds if the hardlockup was disabled but the softlockup
 585          * enabled.
 586          */
 587         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 588                 watchdog_nmi_disable(cpu);
 589 }
 590
 591 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 592 /*
 593  * People like the simple clean cpu node info on boot.
 594  * Reduce the watchdog noise by only printing messages
 595  * that are different from what cpu0 displayed.
 596  */
 597 static unsigned long cpu0_err;
 598
 599 static int watchdog_nmi_enable(unsigned int cpu)
 600 {
 601         struct perf_event_attr *wd_attr;
 602         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 603
 604         /* nothing to do if the hard lockup detector is disabled */
 605         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 606                 goto out;
 607
 608         /* is it already setup and enabled? */
 609         if (event && event->state > PERF_EVENT_STATE_OFF)
 610                 goto out;
 611
 612         /* it is setup but not enabled */
 613         if (event != NULL)
 614                 goto out_enable;
 615
 616         wd_attr = &wd_hw_attr;
 617         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 618
 619         /* Try to register using hardware perf events */
 620         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 621
 622         /* save cpu0 error for future comparision */
 623         if (cpu == 0 && IS_ERR(event))
 624                 cpu0_err = PTR_ERR(event);
 625
 626         if (!IS_ERR(event)) {
 627                 /* only print for cpu0 or different than cpu0 */
 628                 if (cpu == 0 || cpu0_err)
 629                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 630                 goto out_save;
 631         }
 632
 633         /*
 634          * Disable the hard lockup detector if _any_ CPU fails to set up
 635          * set up the hardware perf event. The watchdog() function checks
 636          * the NMI_WATCHDOG_ENABLED bit periodically.
 637          *
 638          * The barriers are for syncing up watchdog_enabled across all the
 639          * cpus, as clear_bit() does not use barriers.
 640          */
 641         smp_mb__before_atomic();
 642         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 643         smp_mb__after_atomic();
 644
 645         /* skip displaying the same error again */
 646         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 647                 return PTR_ERR(event);
 648
 649         /* vary the KERN level based on the returned errno */
 650         if (PTR_ERR(event) == -EOPNOTSUPP)
 651                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 652         else if (PTR_ERR(event) == -ENOENT)
 653                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 654                          cpu);
 655         else
 656                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 657                         cpu, PTR_ERR(event));
 658
 659         pr_info("Shutting down hard lockup detector on all cpus\n");
 660
 661         return PTR_ERR(event);
 662
 663         /* success path */
 664 out_save:
 665         per_cpu(watchdog_ev, cpu) = event;
 666 out_enable:
 667         perf_event_enable(per_cpu(watchdog_ev, cpu));
 668 out:
 669         return 0;
 670 }
 671
 672 static void watchdog_nmi_disable(unsigned int cpu)
 673 {
 674         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 675
 676         if (event) {
 677                 perf_event_disable(event);
 678                 per_cpu(watchdog_ev, cpu) = NULL;
 679
 680                 /* should be in cleanup, but blocks oprofile */
 681                 perf_event_release_kernel(event);
 682         }
 683         if (cpu == 0) {
 684                 /* watchdog_nmi_enable() expects this to be zero initially. */
 685                 cpu0_err = 0;
 686         }
 687 }
 688
 689 #else
 690 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 691 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 692 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 693
 694 static struct smp_hotplug_thread watchdog_threads = {
 695         .store                  = &softlockup_watchdog,
 696         .thread_should_run      = watchdog_should_run,
 697         .thread_fn              = watchdog,
 698         .thread_comm            = "watchdog/%u",
 699         .setup                  = watchdog_enable,
 700         .cleanup                = watchdog_cleanup,
 701         .park                   = watchdog_disable,
 702         .unpark                 = watchdog_enable,
 703 };
 704
 705 /*
 706  * park all watchdog threads that are specified in 'watchdog_cpumask'
 707  *
 708  * This function returns an error if kthread_park() of a watchdog thread
 709  * fails. In this situation, the watchdog threads of some CPUs can already
 710  * be parked and the watchdog threads of other CPUs can still be runnable.
 711  * Callers are expected to handle this special condition as appropriate in
 712  * their context.
 713  *
 714  * This function may only be called in a context that is protected against
 715  * races with CPU hotplug - for example, via get_online_cpus().
 716  */
 717 static int watchdog_park_threads(void)
 718 {
 719         int cpu, ret = 0;
 720
 721         for_each_watchdog_cpu(cpu) {
 722                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 723                 if (ret)
 724                         break;
 725         }
 726
 727         return ret;
 728 }
 729
 730 /*
 731  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 732  *
 733  * This function may only be called in a context that is protected against
 734  * races with CPU hotplug - for example, via get_online_cpus().
 735  */
 736 static void watchdog_unpark_threads(void)
 737 {
 738         int cpu;
 739
 740         for_each_watchdog_cpu(cpu)
 741                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 742 }
 743
 744 /*
 745  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 746  */
 747 int lockup_detector_suspend(void)
 748 {
 749         int ret = 0;
 750
 751         get_online_cpus();
 752         mutex_lock(&watchdog_proc_mutex);
 753         /*
 754          * Multiple suspend requests can be active in parallel (counted by
 755          * the 'watchdog_suspended' variable). If the watchdog threads are
 756          * running, the first caller takes care that they will be parked.
 757          * The state of 'watchdog_running' cannot change while a suspend
 758          * request is active (see related code in 'proc' handlers).
 759          */
 760         if (watchdog_running && !watchdog_suspended)
 761                 ret = watchdog_park_threads();
 762
 763         if (ret == 0)
 764                 watchdog_suspended++;
 765         else {
 766                 watchdog_disable_all_cpus();
 767                 pr_err("Failed to suspend lockup detectors, disabled\n");
 768                 watchdog_enabled = 0;
 769         }
 770
 771         mutex_unlock(&watchdog_proc_mutex);
 772
 773         return ret;
 774 }
 775
 776 /*
 777  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 778  */
 779 void lockup_detector_resume(void)
 780 {
 781         mutex_lock(&watchdog_proc_mutex);
 782
 783         watchdog_suspended--;
 784         /*
 785          * The watchdog threads are unparked if they were previously running
 786          * and if there is no more active suspend request.
 787          */
 788         if (watchdog_running && !watchdog_suspended)
 789                 watchdog_unpark_threads();
 790
 791         mutex_unlock(&watchdog_proc_mutex);
 792         put_online_cpus();
 793 }
 794
 795 static int update_watchdog_all_cpus(void)
 796 {
 797         int ret;
 798
 799         ret = watchdog_park_threads();
 800         if (ret)
 801                 return ret;
 802
 803         watchdog_unpark_threads();
 804
 805         return 0;
 806 }
 807
 808 static int watchdog_enable_all_cpus(void)
 809 {
 810         int err = 0;
 811
 812         if (!watchdog_running) {
 813                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 814                                                              &watchdog_cpumask);
 815                 if (err)
 816                         pr_err("Failed to create watchdog threads, disabled\n");
 817                 else
 818                         watchdog_running = 1;
 819         } else {
 820                 /*
 821                  * Enable/disable the lockup detectors or
 822                  * change the sample period 'on the fly'.
 823                  */
 824                 err = update_watchdog_all_cpus();
 825
 826                 if (err) {
 827                         watchdog_disable_all_cpus();
 828                         pr_err("Failed to update lockup detectors, disabled\n");
 829                 }
 830         }
 831
 832         if (err)
 833                 watchdog_enabled = 0;
 834
 835         return err;
 836 }
 837
 838 static void watchdog_disable_all_cpus(void)
 839 {
 840         if (watchdog_running) {
 841                 watchdog_running = 0;
 842                 smpboot_unregister_percpu_thread(&watchdog_threads);
 843         }
 844 }
 845
 846 #ifdef CONFIG_SYSCTL
 847
 848 /*
 849  * Update the run state of the lockup detectors.
 850  */
 851 static int proc_watchdog_update(void)
 852 {
 853         int err = 0;
 854
 855         /*
 856          * Watchdog threads won't be started if they are already active.
 857          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 858          * care of this. If those threads are already active, the sample
 859          * period will be updated and the lockup detectors will be enabled
 860          * or disabled 'on the fly'.
 861          */
 862         if (watchdog_enabled && watchdog_thresh)
 863                 err = watchdog_enable_all_cpus();
 864         else
 865                 watchdog_disable_all_cpus();
 866
 867         return err;
 868
 869 }
 870
 871 /*
 872  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 873  *
 874  * caller             | table->data points to | 'which' contains the flag(s)
 875  * -------------------|-----------------------|-----------------------------
 876  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 877  *                    |                       | with SOFT_WATCHDOG_ENABLED
 878  * -------------------|-----------------------|-----------------------------
 879  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 880  * -------------------|-----------------------|-----------------------------
 881  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 882  */
 883 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 884                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 885 {
 886         int err, old, new;
 887         int *watchdog_param = (int *)table->data;
 888
 889         get_online_cpus();
 890         mutex_lock(&watchdog_proc_mutex);
 891
 892         if (watchdog_suspended) {
 893                 /* no parameter changes allowed while watchdog is suspended */
 894                 err = -EAGAIN;
 895                 goto out;
 896         }
 897
 898         /*
 899          * If the parameter is being read return the state of the corresponding
 900          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 901          * run state of the lockup detectors.
 902          */
 903         if (!write) {
 904                 *watchdog_param = (watchdog_enabled & which) != 0;
 905                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 906         } else {
 907                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 908                 if (err)
 909                         goto out;
 910
 911                 /*
 912                  * There is a race window between fetching the current value
 913                  * from 'watchdog_enabled' and storing the new value. During
 914                  * this race window, watchdog_nmi_enable() can sneak in and
 915                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 916                  * The 'cmpxchg' detects this race and the loop retries.
 917                  */
 918                 do {
 919                         old = watchdog_enabled;
 920                         /*
 921                          * If the parameter value is not zero set the
 922                          * corresponding bit(s), else clear it(them).
 923                          */
 924                         if (*watchdog_param)
 925                                 new = old | which;
 926                         else
 927                                 new = old & ~which;
 928                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 929
 930                 /*
 931                  * Update the run state of the lockup detectors. There is _no_
 932                  * need to check the value returned by proc_watchdog_update()
 933                  * and to restore the previous value of 'watchdog_enabled' as
 934                  * both lockup detectors are disabled if proc_watchdog_update()
 935                  * returns an error.
 936                  */
 937                 if (old == new)
 938                         goto out;
 939
 940                 err = proc_watchdog_update();
 941         }
 942 out:
 943         mutex_unlock(&watchdog_proc_mutex);
 944         put_online_cpus();
 945         return err;
 946 }
 947
 948 /*
 949  * /proc/sys/kernel/watchdog
 950  */
 951 int proc_watchdog(struct ctl_table *table, int write,
 952                   void __user *buffer, size_t *lenp, loff_t *ppos)
 953 {
 954         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 955                                     table, write, buffer, lenp, ppos);
 956 }
 957
 958 /*
 959  * /proc/sys/kernel/nmi_watchdog
 960  */
 961 int proc_nmi_watchdog(struct ctl_table *table, int write,
 962                       void __user *buffer, size_t *lenp, loff_t *ppos)
 963 {
 964         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 965                                     table, write, buffer, lenp, ppos);
 966 }
 967
 968 /*
 969  * /proc/sys/kernel/soft_watchdog
 970  */
 971 int proc_soft_watchdog(struct ctl_table *table, int write,
 972                         void __user *buffer, size_t *lenp, loff_t *ppos)
 973 {
 974         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 975                                     table, write, buffer, lenp, ppos);
 976 }
 977
 978 /*
 979  * /proc/sys/kernel/watchdog_thresh
 980  */
 981 int proc_watchdog_thresh(struct ctl_table *table, int write,
 982                          void __user *buffer, size_t *lenp, loff_t *ppos)
 983 {
 984         int err, old, new;
 985
 986         get_online_cpus();
 987         mutex_lock(&watchdog_proc_mutex);
 988
 989         if (watchdog_suspended) {
 990                 /* no parameter changes allowed while watchdog is suspended */
 991                 err = -EAGAIN;
 992                 goto out;
 993         }
 994
 995         old = ACCESS_ONCE(watchdog_thresh);
 996         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 997
 998         if (err || !write)
 999                 goto out;
1000
1001         /*
1002          * Update the sample period. Restore on failure.
1003          */
1004         new = ACCESS_ONCE(watchdog_thresh);
1005         if (old == new)
1006                 goto out;
1007
1008         set_sample_period();
1009         err = proc_watchdog_update();
1010         if (err) {
1011                 watchdog_thresh = old;
1012                 set_sample_period();
1013         }
1014 out:
1015         mutex_unlock(&watchdog_proc_mutex);
1016         put_online_cpus();
1017         return err;
1018 }
1019
1020 /*
1021  * The cpumask is the mask of possible cpus that the watchdog can run
1022  * on, not the mask of cpus it is actually running on.  This allows the
1023  * user to specify a mask that will include cpus that have not yet
1024  * been brought online, if desired.
1025  */
1026 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1027                           void __user *buffer, size_t *lenp, loff_t *ppos)
1028 {
1029         int err;
1030
1031         get_online_cpus();
1032         mutex_lock(&watchdog_proc_mutex);
1033
1034         if (watchdog_suspended) {
1035                 /* no parameter changes allowed while watchdog is suspended */
1036                 err = -EAGAIN;
1037                 goto out;
1038         }
1039
1040         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1041         if (!err && write) {
1042                 /* Remove impossible cpus to keep sysctl output cleaner. */
1043                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1044                             cpu_possible_mask);
1045
1046                 if (watchdog_running) {
1047                         /*
1048                          * Failure would be due to being unable to allocate
1049                          * a temporary cpumask, so we are likely not in a
1050                          * position to do much else to make things better.
1051                          */
1052                         if (smpboot_update_cpumask_percpu_thread(
1053                                     &watchdog_threads, &watchdog_cpumask) != 0)
1054                                 pr_err("cpumask update failed\n");
1055                 }
1056         }
1057 out:
1058         mutex_unlock(&watchdog_proc_mutex);
1059         put_online_cpus();
1060         return err;
1061 }
1062
1063 #endif /* CONFIG_SYSCTL */
1064
1065 void __init lockup_detector_init(void)
1066 {
1067         set_sample_period();
1068
1069 #ifdef CONFIG_NO_HZ_FULL
1070         if (tick_nohz_full_enabled()) {
1071                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1072                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1073         } else
1074                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1075 #else
1076         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1077 #endif
1078
1079         if (watchdog_enabled)
1080                 watchdog_enable_all_cpus();
1081 }