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 #endif
 107 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
 108 static cpumask_t __read_mostly watchdog_cpus;
 109 #endif
 110 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
 111 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
 112 #endif
 113 static unsigned long soft_lockup_nmi_warn;
 114
 115 /* boot commands */
 116 /*
 117  * Should we panic when a soft-lockup or hard-lockup occurs:
 118  */
 119 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 120 unsigned int __read_mostly hardlockup_panic =
 121                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
 122 static unsigned long __maybe_unused hardlockup_allcpu_dumped;
 123 /*
 124  * We may not want to enable hard lockup detection by default in all cases,
 125  * for example when running the kernel as a guest on a hypervisor. In these
 126  * cases this function can be called to disable hard lockup detection. This
 127  * function should only be executed once by the boot processor before the
 128  * kernel command line parameters are parsed, because otherwise it is not
 129  * possible to override this in hardlockup_panic_setup().
 130  */
 131 void hardlockup_detector_disable(void)
 132 {
 133         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 134 }
 135
 136 static int __init hardlockup_panic_setup(char *str)
 137 {
 138         if (!strncmp(str, "panic", 5))
 139                 hardlockup_panic = 1;
 140         else if (!strncmp(str, "nopanic", 7))
 141                 hardlockup_panic = 0;
 142         else if (!strncmp(str, "0", 1))
 143                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
 144         else if (!strncmp(str, "1", 1))
 145                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
 146         return 1;
 147 }
 148 __setup("nmi_watchdog=", hardlockup_panic_setup);
 149 #endif
 150
 151 unsigned int __read_mostly softlockup_panic =
 152                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 153
 154 static int __init softlockup_panic_setup(char *str)
 155 {
 156         softlockup_panic = simple_strtoul(str, NULL, 0);
 157
 158         return 1;
 159 }
 160 __setup("softlockup_panic=", softlockup_panic_setup);
 161
 162 static int __init nowatchdog_setup(char *str)
 163 {
 164         watchdog_enabled = 0;
 165         return 1;
 166 }
 167 __setup("nowatchdog", nowatchdog_setup);
 168
 169 static int __init nosoftlockup_setup(char *str)
 170 {
 171         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 172         return 1;
 173 }
 174 __setup("nosoftlockup", nosoftlockup_setup);
 175
 176 #ifdef CONFIG_SMP
 177 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 178 {
 179         sysctl_softlockup_all_cpu_backtrace =
 180                 !!simple_strtol(str, NULL, 0);
 181         return 1;
 182 }
 183 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 184 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
 185 {
 186         sysctl_hardlockup_all_cpu_backtrace =
 187                 !!simple_strtol(str, NULL, 0);
 188         return 1;
 189 }
 190 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
 191 #endif
 192
 193 /*
 194  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 195  * lockups can have false positives under extreme conditions. So we generally
 196  * want a higher threshold for soft lockups than for hard lockups. So we couple
 197  * the thresholds with a factor: we make the soft threshold twice the amount of
 198  * time the hard threshold is.
 199  */
 200 static int get_softlockup_thresh(void)
 201 {
 202         return watchdog_thresh * 2;
 203 }
 204
 205 /*
 206  * Returns seconds, approximately.  We don't need nanosecond
 207  * resolution, and we don't need to waste time with a big divide when
 208  * 2^30ns == 1.074s.
 209  */
 210 static unsigned long get_timestamp(void)
 211 {
 212         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 213 }
 214
 215 static void set_sample_period(void)
 216 {
 217         /*
 218          * convert watchdog_thresh from seconds to ns
 219          * the divide by 5 is to give hrtimer several chances (two
 220          * or three with the current relation between the soft
 221          * and hard thresholds) to increment before the
 222          * hardlockup detector generates a warning
 223          */
 224         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 225 }
 226
 227 /* Commands for resetting the watchdog */
 228 static void __touch_watchdog(void)
 229 {
 230         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 231 }
 232
 233 void touch_softlockup_watchdog(void)
 234 {
 235         /*
 236          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 237          * gets zeroed here, so use the raw_ operation.
 238          */
 239         raw_cpu_write(watchdog_touch_ts, 0);
 240 }
 241 EXPORT_SYMBOL(touch_softlockup_watchdog);
 242
 243 void touch_all_softlockup_watchdogs(void)
 244 {
 245         int cpu;
 246
 247         /*
 248          * this is done lockless
 249          * do we care if a 0 races with a timestamp?
 250          * all it means is the softlock check starts one cycle later
 251          */
 252         for_each_watchdog_cpu(cpu)
 253                 per_cpu(watchdog_touch_ts, cpu) = 0;
 254 }
 255
 256 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 257 void touch_nmi_watchdog(void)
 258 {
 259         /*
 260          * Using __raw here because some code paths have
 261          * preemption enabled.  If preemption is enabled
 262          * then interrupts should be enabled too, in which
 263          * case we shouldn't have to worry about the watchdog
 264          * going off.
 265          */
 266         raw_cpu_write(watchdog_nmi_touch, true);
 267         touch_softlockup_watchdog();
 268 }
 269 EXPORT_SYMBOL(touch_nmi_watchdog);
 270
 271 #endif
 272
 273 void touch_softlockup_watchdog_sync(void)
 274 {
 275         __this_cpu_write(softlockup_touch_sync, true);
 276         __this_cpu_write(watchdog_touch_ts, 0);
 277 }
 278
 279 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
 280 /* watchdog detector functions */
 281 static bool is_hardlockup(void)
 282 {
 283         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 284
 285         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 286                 return true;
 287
 288         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 289         return false;
 290 }
 291 #endif
 292
 293 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
 294 static unsigned int watchdog_next_cpu(unsigned int cpu)
 295 {
 296         cpumask_t cpus = watchdog_cpus;
 297         unsigned int next_cpu;
 298
 299         next_cpu = cpumask_next(cpu, &cpus);
 300         if (next_cpu >= nr_cpu_ids)
 301                 next_cpu = cpumask_first(&cpus);
 302
 303         if (next_cpu == cpu)
 304                 return nr_cpu_ids;
 305
 306         return next_cpu;
 307 }
 308
 309 static int is_hardlockup_other_cpu(unsigned int cpu)
 310 {
 311         unsigned long hrint = per_cpu(hrtimer_interrupts, cpu);
 312
 313         if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
 314                 return 1;
 315
 316         per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
 317         return 0;
 318 }
 319
 320 static void watchdog_check_hardlockup_other_cpu(void)
 321 {
 322         unsigned int next_cpu;
 323
 324         /*
 325          * Test for hardlockups every 3 samples.  The sample period is
 326          *  watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over
 327          *  watchdog_thresh (over by 20%).
 328          */
 329         if (__this_cpu_read(hrtimer_interrupts) % 3 != 0)
 330                 return;
 331
 332         /* check for a hardlockup on the next cpu */
 333         next_cpu = watchdog_next_cpu(smp_processor_id());
 334         if (next_cpu >= nr_cpu_ids)
 335                 return;
 336
 337         smp_rmb();
 338
 339         if (per_cpu(watchdog_nmi_touch, next_cpu) == true) {
 340                 per_cpu(watchdog_nmi_touch, next_cpu) = false;
 341                 return;
 342         }
 343
 344         if (is_hardlockup_other_cpu(next_cpu)) {
 345                 /* only warn once */
 346                 if (per_cpu(hard_watchdog_warn, next_cpu) == true)
 347                         return;
 348
 349                 if (hardlockup_panic)
 350                         panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu);
 351                 else
 352                         WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu);
 353
 354                 per_cpu(hard_watchdog_warn, next_cpu) = true;
 355         } else {
 356                 per_cpu(hard_watchdog_warn, next_cpu) = false;
 357         }
 358 }
 359 #else
 360 static inline void watchdog_check_hardlockup_other_cpu(void) { return; }
 361 #endif
 362
 363 static int is_softlockup(unsigned long touch_ts)
 364 {
 365         unsigned long now = get_timestamp();
 366
 367         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 368                 /* Warn about unreasonable delays. */
 369                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 370                         return now - touch_ts;
 371         }
 372         return 0;
 373 }
 374
 375 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
 376
 377 static struct perf_event_attr wd_hw_attr = {
 378         .type           = PERF_TYPE_HARDWARE,
 379         .config         = PERF_COUNT_HW_CPU_CYCLES,
 380         .size           = sizeof(struct perf_event_attr),
 381         .pinned         = 1,
 382         .disabled       = 1,
 383 };
 384
 385 /* Callback function for perf event subsystem */
 386 static void watchdog_overflow_callback(struct perf_event *event,
 387                  struct perf_sample_data *data,
 388                  struct pt_regs *regs)
 389 {
 390         /* Ensure the watchdog never gets throttled */
 391         event->hw.interrupts = 0;
 392
 393         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 394                 __this_cpu_write(watchdog_nmi_touch, false);
 395                 return;
 396         }
 397
 398         /* check for a hardlockup
 399          * This is done by making sure our timer interrupt
 400          * is incrementing.  The timer interrupt should have
 401          * fired multiple times before we overflow'd.  If it hasn't
 402          * then this is a good indication the cpu is stuck
 403          */
 404         if (is_hardlockup()) {
 405                 int this_cpu = smp_processor_id();
 406                 struct pt_regs *regs = get_irq_regs();
 407
 408                 /* only print hardlockups once */
 409                 if (__this_cpu_read(hard_watchdog_warn) == true)
 410                         return;
 411
 412                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 413                 print_modules();
 414                 print_irqtrace_events(current);
 415                 if (regs)
 416                         show_regs(regs);
 417                 else
 418                         dump_stack();
 419
 420                 /*
 421                  * Perform all-CPU dump only once to avoid multiple hardlockups
 422                  * generating interleaving traces
 423                  */
 424                 if (sysctl_hardlockup_all_cpu_backtrace &&
 425                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 426                         trigger_allbutself_cpu_backtrace();
 427
 428                 if (hardlockup_panic)
 429                         panic("Hard LOCKUP");
 430
 431                 __this_cpu_write(hard_watchdog_warn, true);
 432                 return;
 433         }
 434
 435         __this_cpu_write(hard_watchdog_warn, false);
 436         return;
 437 }
 438 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
 439
 440 static void watchdog_interrupt_count(void)
 441 {
 442         __this_cpu_inc(hrtimer_interrupts);
 443 }
 444
 445 static int watchdog_nmi_enable(unsigned int cpu);
 446 static void watchdog_nmi_disable(unsigned int cpu);
 447
 448 static int watchdog_enable_all_cpus(void);
 449 static void watchdog_disable_all_cpus(void);
 450
 451 /* watchdog kicker functions */
 452 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 453 {
 454         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 455         struct pt_regs *regs = get_irq_regs();
 456         int duration;
 457         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 458
 459         /* kick the hardlockup detector */
 460         watchdog_interrupt_count();
 461
 462         /* test for hardlockups on the next cpu */
 463         watchdog_check_hardlockup_other_cpu();
 464
 465         /* kick the softlockup detector */
 466         wake_up_process(__this_cpu_read(softlockup_watchdog));
 467
 468         /* .. and repeat */
 469         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 470
 471         if (touch_ts == 0) {
 472                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 473                         /*
 474                          * If the time stamp was touched atomically
 475                          * make sure the scheduler tick is up to date.
 476                          */
 477                         __this_cpu_write(softlockup_touch_sync, false);
 478                         sched_clock_tick();
 479                 }
 480
 481                 /* Clear the guest paused flag on watchdog reset */
 482                 kvm_check_and_clear_guest_paused();
 483                 __touch_watchdog();
 484                 return HRTIMER_RESTART;
 485         }
 486
 487         /* check for a softlockup
 488          * This is done by making sure a high priority task is
 489          * being scheduled.  The task touches the watchdog to
 490          * indicate it is getting cpu time.  If it hasn't then
 491          * this is a good indication some task is hogging the cpu
 492          */
 493         duration = is_softlockup(touch_ts);
 494         if (unlikely(duration)) {
 495                 /*
 496                  * If a virtual machine is stopped by the host it can look to
 497                  * the watchdog like a soft lockup, check to see if the host
 498                  * stopped the vm before we issue the warning
 499                  */
 500                 if (kvm_check_and_clear_guest_paused())
 501                         return HRTIMER_RESTART;
 502
 503                 /* only warn once */
 504                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 505                         /*
 506                          * When multiple processes are causing softlockups the
 507                          * softlockup detector only warns on the first one
 508                          * because the code relies on a full quiet cycle to
 509                          * re-arm.  The second process prevents the quiet cycle
 510                          * and never gets reported.  Use task pointers to detect
 511                          * this.
 512                          */
 513                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 514                             current) {
 515                                 __this_cpu_write(soft_watchdog_warn, false);
 516                                 __touch_watchdog();
 517                         }
 518                         return HRTIMER_RESTART;
 519                 }
 520
 521                 if (softlockup_all_cpu_backtrace) {
 522                         /* Prevent multiple soft-lockup reports if one cpu is already
 523                          * engaged in dumping cpu back traces
 524                          */
 525                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 526                                 /* Someone else will report us. Let's give up */
 527                                 __this_cpu_write(soft_watchdog_warn, true);
 528                                 return HRTIMER_RESTART;
 529                         }
 530                 }
 531
 532                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 533                         smp_processor_id(), duration,
 534                         current->comm, task_pid_nr(current));
 535                 __this_cpu_write(softlockup_task_ptr_saved, current);
 536                 print_modules();
 537                 print_irqtrace_events(current);
 538                 if (regs)
 539                         show_regs(regs);
 540                 else
 541                         dump_stack();
 542
 543                 if (softlockup_all_cpu_backtrace) {
 544                         /* Avoid generating two back traces for current
 545                          * given that one is already made above
 546                          */
 547                         trigger_allbutself_cpu_backtrace();
 548
 549                         clear_bit(0, &soft_lockup_nmi_warn);
 550                         /* Barrier to sync with other cpus */
 551                         smp_mb__after_atomic();
 552                 }
 553
 554                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 555                 if (softlockup_panic)
 556                         panic("softlockup: hung tasks");
 557                 __this_cpu_write(soft_watchdog_warn, true);
 558         } else
 559                 __this_cpu_write(soft_watchdog_warn, false);
 560
 561         return HRTIMER_RESTART;
 562 }
 563
 564 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 565 {
 566         struct sched_param param = { .sched_priority = prio };
 567
 568         sched_setscheduler(current, policy, &param);
 569 }
 570
 571 static void watchdog_enable(unsigned int cpu)
 572 {
 573         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 574
 575         /* kick off the timer for the hardlockup detector */
 576         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 577         hrtimer->function = watchdog_timer_fn;
 578
 579         /* Enable the perf event */
 580         watchdog_nmi_enable(cpu);
 581
 582         /* done here because hrtimer_start can only pin to smp_processor_id() */
 583         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 584                       HRTIMER_MODE_REL_PINNED);
 585
 586         /* initialize timestamp */
 587         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 588         __touch_watchdog();
 589 }
 590
 591 static void watchdog_disable(unsigned int cpu)
 592 {
 593         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 594
 595         watchdog_set_prio(SCHED_NORMAL, 0);
 596         hrtimer_cancel(hrtimer);
 597         /* disable the perf event */
 598         watchdog_nmi_disable(cpu);
 599 }
 600
 601 static void watchdog_cleanup(unsigned int cpu, bool online)
 602 {
 603         watchdog_disable(cpu);
 604 }
 605
 606 static int watchdog_should_run(unsigned int cpu)
 607 {
 608         return __this_cpu_read(hrtimer_interrupts) !=
 609                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 610 }
 611
 612 /*
 613  * The watchdog thread function - touches the timestamp.
 614  *
 615  * It only runs once every sample_period seconds (4 seconds by
 616  * default) to reset the softlockup timestamp. If this gets delayed
 617  * for more than 2*watchdog_thresh seconds then the debug-printout
 618  * triggers in watchdog_timer_fn().
 619  */
 620 static void watchdog(unsigned int cpu)
 621 {
 622         __this_cpu_write(soft_lockup_hrtimer_cnt,
 623                          __this_cpu_read(hrtimer_interrupts));
 624         __touch_watchdog();
 625
 626         /*
 627          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 628          * failure path. Check for failures that can occur asynchronously -
 629          * for example, when CPUs are on-lined - and shut down the hardware
 630          * perf event on each CPU accordingly.
 631          *
 632          * The only non-obvious place this bit can be cleared is through
 633          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 634          * pr_info here would be too noisy as it would result in a message
 635          * every few seconds if the hardlockup was disabled but the softlockup
 636          * enabled.
 637          */
 638         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 639                 watchdog_nmi_disable(cpu);
 640 }
 641
 642 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
 643 /*
 644  * People like the simple clean cpu node info on boot.
 645  * Reduce the watchdog noise by only printing messages
 646  * that are different from what cpu0 displayed.
 647  */
 648 static unsigned long cpu0_err;
 649
 650 static int watchdog_nmi_enable(unsigned int cpu)
 651 {
 652         struct perf_event_attr *wd_attr;
 653         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 654
 655         /* nothing to do if the hard lockup detector is disabled */
 656         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 657                 goto out;
 658
 659         /* is it already setup and enabled? */
 660         if (event && event->state > PERF_EVENT_STATE_OFF)
 661                 goto out;
 662
 663         /* it is setup but not enabled */
 664         if (event != NULL)
 665                 goto out_enable;
 666
 667         wd_attr = &wd_hw_attr;
 668         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 669
 670         /* Try to register using hardware perf events */
 671         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 672
 673         /* save cpu0 error for future comparision */
 674         if (cpu == 0 && IS_ERR(event))
 675                 cpu0_err = PTR_ERR(event);
 676
 677         if (!IS_ERR(event)) {
 678                 /* only print for cpu0 or different than cpu0 */
 679                 if (cpu == 0 || cpu0_err)
 680                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 681                 goto out_save;
 682         }
 683
 684         /*
 685          * Disable the hard lockup detector if _any_ CPU fails to set up
 686          * set up the hardware perf event. The watchdog() function checks
 687          * the NMI_WATCHDOG_ENABLED bit periodically.
 688          *
 689          * The barriers are for syncing up watchdog_enabled across all the
 690          * cpus, as clear_bit() does not use barriers.
 691          */
 692         smp_mb__before_atomic();
 693         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 694         smp_mb__after_atomic();
 695
 696         /* skip displaying the same error again */
 697         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 698                 return PTR_ERR(event);
 699
 700         /* vary the KERN level based on the returned errno */
 701         if (PTR_ERR(event) == -EOPNOTSUPP)
 702                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 703         else if (PTR_ERR(event) == -ENOENT)
 704                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 705                          cpu);
 706         else
 707                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 708                         cpu, PTR_ERR(event));
 709
 710         pr_info("Shutting down hard lockup detector on all cpus\n");
 711
 712         return PTR_ERR(event);
 713
 714         /* success path */
 715 out_save:
 716         per_cpu(watchdog_ev, cpu) = event;
 717 out_enable:
 718         perf_event_enable(per_cpu(watchdog_ev, cpu));
 719 out:
 720         return 0;
 721 }
 722
 723 static void watchdog_nmi_disable(unsigned int cpu)
 724 {
 725         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 726
 727         if (event) {
 728                 perf_event_disable(event);
 729                 per_cpu(watchdog_ev, cpu) = NULL;
 730
 731                 /* should be in cleanup, but blocks oprofile */
 732                 perf_event_release_kernel(event);
 733         }
 734         if (cpu == 0) {
 735                 /* watchdog_nmi_enable() expects this to be zero initially. */
 736                 cpu0_err = 0;
 737         }
 738 }
 739
 740 #else
 741 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
 742 static int watchdog_nmi_enable(unsigned int cpu)
 743 {
 744         /*
 745          * The new cpu will be marked online before the first hrtimer interrupt
 746          * runs on it.  If another cpu tests for a hardlockup on the new cpu
 747          * before it has run its first hrtimer, it will get a false positive.
 748          * Touch the watchdog on the new cpu to delay the first check for at
 749          * least 3 sampling periods to guarantee one hrtimer has run on the new
 750          * cpu.
 751          */
 752         per_cpu(watchdog_nmi_touch, cpu) = true;
 753         smp_wmb();
 754         cpumask_set_cpu(cpu, &watchdog_cpus);
 755         return 0;
 756 }
 757
 758 static void watchdog_nmi_disable(unsigned int cpu)
 759 {
 760         unsigned int next_cpu = watchdog_next_cpu(cpu);
 761
 762         /*
 763          * Offlining this cpu will cause the cpu before this one to start
 764          * checking the one after this one.  If this cpu just finished checking
 765          * the next cpu and updating hrtimer_interrupts_saved, and then the
 766          * previous cpu checks it within one sample period, it will trigger a
 767          * false positive.  Touch the watchdog on the next cpu to prevent it.
 768          */
 769         if (next_cpu < nr_cpu_ids)
 770                 per_cpu(watchdog_nmi_touch, next_cpu) = true;
 771         smp_wmb();
 772         cpumask_clear_cpu(cpu, &watchdog_cpus);
 773 }
 774 #else
 775 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 776 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 777 #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */
 778 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
 779
 780 static struct smp_hotplug_thread watchdog_threads = {
 781         .store                  = &softlockup_watchdog,
 782         .thread_should_run      = watchdog_should_run,
 783         .thread_fn              = watchdog,
 784         .thread_comm            = "watchdog/%u",
 785         .setup                  = watchdog_enable,
 786         .cleanup                = watchdog_cleanup,
 787         .park                   = watchdog_disable,
 788         .unpark                 = watchdog_enable,
 789 };
 790
 791 /*
 792  * park all watchdog threads that are specified in 'watchdog_cpumask'
 793  *
 794  * This function returns an error if kthread_park() of a watchdog thread
 795  * fails. In this situation, the watchdog threads of some CPUs can already
 796  * be parked and the watchdog threads of other CPUs can still be runnable.
 797  * Callers are expected to handle this special condition as appropriate in
 798  * their context.
 799  *
 800  * This function may only be called in a context that is protected against
 801  * races with CPU hotplug - for example, via get_online_cpus().
 802  */
 803 static int watchdog_park_threads(void)
 804 {
 805         int cpu, ret = 0;
 806
 807         for_each_watchdog_cpu(cpu) {
 808                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 809                 if (ret)
 810                         break;
 811         }
 812
 813         return ret;
 814 }
 815
 816 /*
 817  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 818  *
 819  * This function may only be called in a context that is protected against
 820  * races with CPU hotplug - for example, via get_online_cpus().
 821  */
 822 static void watchdog_unpark_threads(void)
 823 {
 824         int cpu;
 825
 826         for_each_watchdog_cpu(cpu)
 827                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 828 }
 829
 830 /*
 831  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 832  */
 833 int lockup_detector_suspend(void)
 834 {
 835         int ret = 0;
 836
 837         get_online_cpus();
 838         mutex_lock(&watchdog_proc_mutex);
 839         /*
 840          * Multiple suspend requests can be active in parallel (counted by
 841          * the 'watchdog_suspended' variable). If the watchdog threads are
 842          * running, the first caller takes care that they will be parked.
 843          * The state of 'watchdog_running' cannot change while a suspend
 844          * request is active (see related code in 'proc' handlers).
 845          */
 846         if (watchdog_running && !watchdog_suspended)
 847                 ret = watchdog_park_threads();
 848
 849         if (ret == 0)
 850                 watchdog_suspended++;
 851         else {
 852                 watchdog_disable_all_cpus();
 853                 pr_err("Failed to suspend lockup detectors, disabled\n");
 854                 watchdog_enabled = 0;
 855         }
 856
 857         mutex_unlock(&watchdog_proc_mutex);
 858
 859         return ret;
 860 }
 861
 862 /*
 863  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 864  */
 865 void lockup_detector_resume(void)
 866 {
 867         mutex_lock(&watchdog_proc_mutex);
 868
 869         watchdog_suspended--;
 870         /*
 871          * The watchdog threads are unparked if they were previously running
 872          * and if there is no more active suspend request.
 873          */
 874         if (watchdog_running && !watchdog_suspended)
 875                 watchdog_unpark_threads();
 876
 877         mutex_unlock(&watchdog_proc_mutex);
 878         put_online_cpus();
 879 }
 880
 881 static int update_watchdog_all_cpus(void)
 882 {
 883         int ret;
 884
 885         ret = watchdog_park_threads();
 886         if (ret)
 887                 return ret;
 888
 889         watchdog_unpark_threads();
 890
 891         return 0;
 892 }
 893
 894 static int watchdog_enable_all_cpus(void)
 895 {
 896         int err = 0;
 897
 898         if (!watchdog_running) {
 899                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 900                                                              &watchdog_cpumask);
 901                 if (err)
 902                         pr_err("Failed to create watchdog threads, disabled\n");
 903                 else
 904                         watchdog_running = 1;
 905         } else {
 906                 /*
 907                  * Enable/disable the lockup detectors or
 908                  * change the sample period 'on the fly'.
 909                  */
 910                 err = update_watchdog_all_cpus();
 911
 912                 if (err) {
 913                         watchdog_disable_all_cpus();
 914                         pr_err("Failed to update lockup detectors, disabled\n");
 915                 }
 916         }
 917
 918         if (err)
 919                 watchdog_enabled = 0;
 920
 921         return err;
 922 }
 923
 924 static void watchdog_disable_all_cpus(void)
 925 {
 926         if (watchdog_running) {
 927                 watchdog_running = 0;
 928                 smpboot_unregister_percpu_thread(&watchdog_threads);
 929         }
 930 }
 931
 932 #ifdef CONFIG_SYSCTL
 933
 934 /*
 935  * Update the run state of the lockup detectors.
 936  */
 937 static int proc_watchdog_update(void)
 938 {
 939         int err = 0;
 940
 941         /*
 942          * Watchdog threads won't be started if they are already active.
 943          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 944          * care of this. If those threads are already active, the sample
 945          * period will be updated and the lockup detectors will be enabled
 946          * or disabled 'on the fly'.
 947          */
 948         if (watchdog_enabled && watchdog_thresh)
 949                 err = watchdog_enable_all_cpus();
 950         else
 951                 watchdog_disable_all_cpus();
 952
 953         return err;
 954
 955 }
 956
 957 /*
 958  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 959  *
 960  * caller             | table->data points to | 'which' contains the flag(s)
 961  * -------------------|-----------------------|-----------------------------
 962  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 963  *                    |                       | with SOFT_WATCHDOG_ENABLED
 964  * -------------------|-----------------------|-----------------------------
 965  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 966  * -------------------|-----------------------|-----------------------------
 967  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 968  */
 969 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 970                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 971 {
 972         int err, old, new;
 973         int *watchdog_param = (int *)table->data;
 974
 975         get_online_cpus();
 976         mutex_lock(&watchdog_proc_mutex);
 977
 978         if (watchdog_suspended) {
 979                 /* no parameter changes allowed while watchdog is suspended */
 980                 err = -EAGAIN;
 981                 goto out;
 982         }
 983
 984         /*
 985          * If the parameter is being read return the state of the corresponding
 986          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 987          * run state of the lockup detectors.
 988          */
 989         if (!write) {
 990                 *watchdog_param = (watchdog_enabled & which) != 0;
 991                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 992         } else {
 993                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 994                 if (err)
 995                         goto out;
 996
 997                 /*
 998                  * There is a race window between fetching the current value
 999                  * from 'watchdog_enabled' and storing the new value. During
1000                  * this race window, watchdog_nmi_enable() can sneak in and
1001                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
1002                  * The 'cmpxchg' detects this race and the loop retries.
1003                  */
1004                 do {
1005                         old = watchdog_enabled;
1006                         /*
1007                          * If the parameter value is not zero set the
1008                          * corresponding bit(s), else clear it(them).
1009                          */
1010                         if (*watchdog_param)
1011                                 new = old | which;
1012                         else
1013                                 new = old & ~which;
1014                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
1015
1016                 /*
1017                  * Update the run state of the lockup detectors. There is _no_
1018                  * need to check the value returned by proc_watchdog_update()
1019                  * and to restore the previous value of 'watchdog_enabled' as
1020                  * both lockup detectors are disabled if proc_watchdog_update()
1021                  * returns an error.
1022                  */
1023                 if (old == new)
1024                         goto out;
1025
1026                 err = proc_watchdog_update();
1027         }
1028 out:
1029         mutex_unlock(&watchdog_proc_mutex);
1030         put_online_cpus();
1031         return err;
1032 }
1033
1034 /*
1035  * /proc/sys/kernel/watchdog
1036  */
1037 int proc_watchdog(struct ctl_table *table, int write,
1038                   void __user *buffer, size_t *lenp, loff_t *ppos)
1039 {
1040         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
1041                                     table, write, buffer, lenp, ppos);
1042 }
1043
1044 /*
1045  * /proc/sys/kernel/nmi_watchdog
1046  */
1047 int proc_nmi_watchdog(struct ctl_table *table, int write,
1048                       void __user *buffer, size_t *lenp, loff_t *ppos)
1049 {
1050         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
1051                                     table, write, buffer, lenp, ppos);
1052 }
1053
1054 /*
1055  * /proc/sys/kernel/soft_watchdog
1056  */
1057 int proc_soft_watchdog(struct ctl_table *table, int write,
1058                         void __user *buffer, size_t *lenp, loff_t *ppos)
1059 {
1060         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
1061                                     table, write, buffer, lenp, ppos);
1062 }
1063
1064 /*
1065  * /proc/sys/kernel/watchdog_thresh
1066  */
1067 int proc_watchdog_thresh(struct ctl_table *table, int write,
1068                          void __user *buffer, size_t *lenp, loff_t *ppos)
1069 {
1070         int err, old, new;
1071
1072         get_online_cpus();
1073         mutex_lock(&watchdog_proc_mutex);
1074
1075         if (watchdog_suspended) {
1076                 /* no parameter changes allowed while watchdog is suspended */
1077                 err = -EAGAIN;
1078                 goto out;
1079         }
1080
1081         old = ACCESS_ONCE(watchdog_thresh);
1082         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1083
1084         if (err || !write)
1085                 goto out;
1086
1087         /*
1088          * Update the sample period. Restore on failure.
1089          */
1090         new = ACCESS_ONCE(watchdog_thresh);
1091         if (old == new)
1092                 goto out;
1093
1094         set_sample_period();
1095         err = proc_watchdog_update();
1096         if (err) {
1097                 watchdog_thresh = old;
1098                 set_sample_period();
1099         }
1100 out:
1101         mutex_unlock(&watchdog_proc_mutex);
1102         put_online_cpus();
1103         return err;
1104 }
1105
1106 /*
1107  * The cpumask is the mask of possible cpus that the watchdog can run
1108  * on, not the mask of cpus it is actually running on.  This allows the
1109  * user to specify a mask that will include cpus that have not yet
1110  * been brought online, if desired.
1111  */
1112 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1113                           void __user *buffer, size_t *lenp, loff_t *ppos)
1114 {
1115         int err;
1116
1117         get_online_cpus();
1118         mutex_lock(&watchdog_proc_mutex);
1119
1120         if (watchdog_suspended) {
1121                 /* no parameter changes allowed while watchdog is suspended */
1122                 err = -EAGAIN;
1123                 goto out;
1124         }
1125
1126         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1127         if (!err && write) {
1128                 /* Remove impossible cpus to keep sysctl output cleaner. */
1129                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1130                             cpu_possible_mask);
1131
1132                 if (watchdog_running) {
1133                         /*
1134                          * Failure would be due to being unable to allocate
1135                          * a temporary cpumask, so we are likely not in a
1136                          * position to do much else to make things better.
1137                          */
1138                         if (smpboot_update_cpumask_percpu_thread(
1139                                     &watchdog_threads, &watchdog_cpumask) != 0)
1140                                 pr_err("cpumask update failed\n");
1141                 }
1142         }
1143 out:
1144         mutex_unlock(&watchdog_proc_mutex);
1145         put_online_cpus();
1146         return err;
1147 }
1148
1149 #endif /* CONFIG_SYSCTL */
1150
1151 void __init lockup_detector_init(void)
1152 {
1153         set_sample_period();
1154
1155 #ifdef CONFIG_NO_HZ_FULL
1156         if (tick_nohz_full_enabled()) {
1157                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1158                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1159         } else
1160                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1161 #else
1162         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1163 #endif
1164
1165         if (watchdog_enabled)
1166                 watchdog_enable_all_cpus();
1167 }