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
 407                 /* only print hardlockups once */
 408                 if (__this_cpu_read(hard_watchdog_warn) == true)
 409                         return;
 410
 411                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 412                 print_modules();
 413                 print_irqtrace_events(current);
 414                 if (regs)
 415                         show_regs(regs);
 416                 else
 417                         dump_stack();
 418
 419                 /*
 420                  * Perform all-CPU dump only once to avoid multiple hardlockups
 421                  * generating interleaving traces
 422                  */
 423                 if (sysctl_hardlockup_all_cpu_backtrace &&
 424                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 425                         trigger_allbutself_cpu_backtrace();
 426
 427                 if (hardlockup_panic)
 428                         panic("Hard LOCKUP");
 429
 430                 __this_cpu_write(hard_watchdog_warn, true);
 431                 return;
 432         }
 433
 434         __this_cpu_write(hard_watchdog_warn, false);
 435         return;
 436 }
 437 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
 438
 439 static void watchdog_interrupt_count(void)
 440 {
 441         __this_cpu_inc(hrtimer_interrupts);
 442 }
 443
 444 static int watchdog_nmi_enable(unsigned int cpu);
 445 static void watchdog_nmi_disable(unsigned int cpu);
 446
 447 static int watchdog_enable_all_cpus(void);
 448 static void watchdog_disable_all_cpus(void);
 449
 450 /* watchdog kicker functions */
 451 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 452 {
 453         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 454         struct pt_regs *regs = get_irq_regs();
 455         int duration;
 456         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 457
 458         /* kick the hardlockup detector */
 459         watchdog_interrupt_count();
 460
 461         /* test for hardlockups on the next cpu */
 462         watchdog_check_hardlockup_other_cpu();
 463
 464         /* kick the softlockup detector */
 465         wake_up_process(__this_cpu_read(softlockup_watchdog));
 466
 467         /* .. and repeat */
 468         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 469
 470         if (touch_ts == 0) {
 471                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 472                         /*
 473                          * If the time stamp was touched atomically
 474                          * make sure the scheduler tick is up to date.
 475                          */
 476                         __this_cpu_write(softlockup_touch_sync, false);
 477                         sched_clock_tick();
 478                 }
 479
 480                 /* Clear the guest paused flag on watchdog reset */
 481                 kvm_check_and_clear_guest_paused();
 482                 __touch_watchdog();
 483                 return HRTIMER_RESTART;
 484         }
 485
 486         /* check for a softlockup
 487          * This is done by making sure a high priority task is
 488          * being scheduled.  The task touches the watchdog to
 489          * indicate it is getting cpu time.  If it hasn't then
 490          * this is a good indication some task is hogging the cpu
 491          */
 492         duration = is_softlockup(touch_ts);
 493         if (unlikely(duration)) {
 494                 /*
 495                  * If a virtual machine is stopped by the host it can look to
 496                  * the watchdog like a soft lockup, check to see if the host
 497                  * stopped the vm before we issue the warning
 498                  */
 499                 if (kvm_check_and_clear_guest_paused())
 500                         return HRTIMER_RESTART;
 501
 502                 /* only warn once */
 503                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 504                         /*
 505                          * When multiple processes are causing softlockups the
 506                          * softlockup detector only warns on the first one
 507                          * because the code relies on a full quiet cycle to
 508                          * re-arm.  The second process prevents the quiet cycle
 509                          * and never gets reported.  Use task pointers to detect
 510                          * this.
 511                          */
 512                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 513                             current) {
 514                                 __this_cpu_write(soft_watchdog_warn, false);
 515                                 __touch_watchdog();
 516                         }
 517                         return HRTIMER_RESTART;
 518                 }
 519
 520                 if (softlockup_all_cpu_backtrace) {
 521                         /* Prevent multiple soft-lockup reports if one cpu is already
 522                          * engaged in dumping cpu back traces
 523                          */
 524                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 525                                 /* Someone else will report us. Let's give up */
 526                                 __this_cpu_write(soft_watchdog_warn, true);
 527                                 return HRTIMER_RESTART;
 528                         }
 529                 }
 530
 531                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 532                         smp_processor_id(), duration,
 533                         current->comm, task_pid_nr(current));
 534                 __this_cpu_write(softlockup_task_ptr_saved, current);
 535                 print_modules();
 536                 print_irqtrace_events(current);
 537                 if (regs)
 538                         show_regs(regs);
 539                 else
 540                         dump_stack();
 541
 542                 if (softlockup_all_cpu_backtrace) {
 543                         /* Avoid generating two back traces for current
 544                          * given that one is already made above
 545                          */
 546                         trigger_allbutself_cpu_backtrace();
 547
 548                         clear_bit(0, &soft_lockup_nmi_warn);
 549                         /* Barrier to sync with other cpus */
 550                         smp_mb__after_atomic();
 551                 }
 552
 553                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 554                 if (softlockup_panic)
 555                         panic("softlockup: hung tasks");
 556                 __this_cpu_write(soft_watchdog_warn, true);
 557         } else
 558                 __this_cpu_write(soft_watchdog_warn, false);
 559
 560         return HRTIMER_RESTART;
 561 }
 562
 563 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 564 {
 565         struct sched_param param = { .sched_priority = prio };
 566
 567         sched_setscheduler(current, policy, &param);
 568 }
 569
 570 static void watchdog_enable(unsigned int cpu)
 571 {
 572         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 573
 574         /* kick off the timer for the hardlockup detector */
 575         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 576         hrtimer->function = watchdog_timer_fn;
 577
 578         /* Enable the perf event */
 579         watchdog_nmi_enable(cpu);
 580
 581         /* done here because hrtimer_start can only pin to smp_processor_id() */
 582         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 583                       HRTIMER_MODE_REL_PINNED);
 584
 585         /* initialize timestamp */
 586         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 587         __touch_watchdog();
 588 }
 589
 590 static void watchdog_disable(unsigned int cpu)
 591 {
 592         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 593
 594         watchdog_set_prio(SCHED_NORMAL, 0);
 595         hrtimer_cancel(hrtimer);
 596         /* disable the perf event */
 597         watchdog_nmi_disable(cpu);
 598 }
 599
 600 static void watchdog_cleanup(unsigned int cpu, bool online)
 601 {
 602         watchdog_disable(cpu);
 603 }
 604
 605 static int watchdog_should_run(unsigned int cpu)
 606 {
 607         return __this_cpu_read(hrtimer_interrupts) !=
 608                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 609 }
 610
 611 /*
 612  * The watchdog thread function - touches the timestamp.
 613  *
 614  * It only runs once every sample_period seconds (4 seconds by
 615  * default) to reset the softlockup timestamp. If this gets delayed
 616  * for more than 2*watchdog_thresh seconds then the debug-printout
 617  * triggers in watchdog_timer_fn().
 618  */
 619 static void watchdog(unsigned int cpu)
 620 {
 621         __this_cpu_write(soft_lockup_hrtimer_cnt,
 622                          __this_cpu_read(hrtimer_interrupts));
 623         __touch_watchdog();
 624
 625         /*
 626          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 627          * failure path. Check for failures that can occur asynchronously -
 628          * for example, when CPUs are on-lined - and shut down the hardware
 629          * perf event on each CPU accordingly.
 630          *
 631          * The only non-obvious place this bit can be cleared is through
 632          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 633          * pr_info here would be too noisy as it would result in a message
 634          * every few seconds if the hardlockup was disabled but the softlockup
 635          * enabled.
 636          */
 637         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 638                 watchdog_nmi_disable(cpu);
 639 }
 640
 641 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
 642 /*
 643  * People like the simple clean cpu node info on boot.
 644  * Reduce the watchdog noise by only printing messages
 645  * that are different from what cpu0 displayed.
 646  */
 647 static unsigned long cpu0_err;
 648
 649 static int watchdog_nmi_enable(unsigned int cpu)
 650 {
 651         struct perf_event_attr *wd_attr;
 652         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 653
 654         /* nothing to do if the hard lockup detector is disabled */
 655         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 656                 goto out;
 657
 658         /* is it already setup and enabled? */
 659         if (event && event->state > PERF_EVENT_STATE_OFF)
 660                 goto out;
 661
 662         /* it is setup but not enabled */
 663         if (event != NULL)
 664                 goto out_enable;
 665
 666         wd_attr = &wd_hw_attr;
 667         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 668
 669         /* Try to register using hardware perf events */
 670         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 671
 672         /* save cpu0 error for future comparision */
 673         if (cpu == 0 && IS_ERR(event))
 674                 cpu0_err = PTR_ERR(event);
 675
 676         if (!IS_ERR(event)) {
 677                 /* only print for cpu0 or different than cpu0 */
 678                 if (cpu == 0 || cpu0_err)
 679                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 680                 goto out_save;
 681         }
 682
 683         /*
 684          * Disable the hard lockup detector if _any_ CPU fails to set up
 685          * set up the hardware perf event. The watchdog() function checks
 686          * the NMI_WATCHDOG_ENABLED bit periodically.
 687          *
 688          * The barriers are for syncing up watchdog_enabled across all the
 689          * cpus, as clear_bit() does not use barriers.
 690          */
 691         smp_mb__before_atomic();
 692         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 693         smp_mb__after_atomic();
 694
 695         /* skip displaying the same error again */
 696         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 697                 return PTR_ERR(event);
 698
 699         /* vary the KERN level based on the returned errno */
 700         if (PTR_ERR(event) == -EOPNOTSUPP)
 701                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 702         else if (PTR_ERR(event) == -ENOENT)
 703                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 704                          cpu);
 705         else
 706                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 707                         cpu, PTR_ERR(event));
 708
 709         pr_info("Shutting down hard lockup detector on all cpus\n");
 710
 711         return PTR_ERR(event);
 712
 713         /* success path */
 714 out_save:
 715         per_cpu(watchdog_ev, cpu) = event;
 716 out_enable:
 717         perf_event_enable(per_cpu(watchdog_ev, cpu));
 718 out:
 719         return 0;
 720 }
 721
 722 static void watchdog_nmi_disable(unsigned int cpu)
 723 {
 724         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 725
 726         if (event) {
 727                 perf_event_disable(event);
 728                 per_cpu(watchdog_ev, cpu) = NULL;
 729
 730                 /* should be in cleanup, but blocks oprofile */
 731                 perf_event_release_kernel(event);
 732         }
 733         if (cpu == 0) {
 734                 /* watchdog_nmi_enable() expects this to be zero initially. */
 735                 cpu0_err = 0;
 736         }
 737 }
 738
 739 #else
 740 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
 741 static int watchdog_nmi_enable(unsigned int cpu)
 742 {
 743         /*
 744          * The new cpu will be marked online before the first hrtimer interrupt
 745          * runs on it.  If another cpu tests for a hardlockup on the new cpu
 746          * before it has run its first hrtimer, it will get a false positive.
 747          * Touch the watchdog on the new cpu to delay the first check for at
 748          * least 3 sampling periods to guarantee one hrtimer has run on the new
 749          * cpu.
 750          */
 751         per_cpu(watchdog_nmi_touch, cpu) = true;
 752         smp_wmb();
 753         cpumask_set_cpu(cpu, &watchdog_cpus);
 754         return 0;
 755 }
 756
 757 static void watchdog_nmi_disable(unsigned int cpu)
 758 {
 759         unsigned int next_cpu = watchdog_next_cpu(cpu);
 760
 761         /*
 762          * Offlining this cpu will cause the cpu before this one to start
 763          * checking the one after this one.  If this cpu just finished checking
 764          * the next cpu and updating hrtimer_interrupts_saved, and then the
 765          * previous cpu checks it within one sample period, it will trigger a
 766          * false positive.  Touch the watchdog on the next cpu to prevent it.
 767          */
 768         if (next_cpu < nr_cpu_ids)
 769                 per_cpu(watchdog_nmi_touch, next_cpu) = true;
 770         smp_wmb();
 771         cpumask_clear_cpu(cpu, &watchdog_cpus);
 772 }
 773 #else
 774 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
 775 static void watchdog_nmi_disable(unsigned int cpu) { return; }
 776 #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */
 777 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
 778
 779 static struct smp_hotplug_thread watchdog_threads = {
 780         .store                  = &softlockup_watchdog,
 781         .thread_should_run      = watchdog_should_run,
 782         .thread_fn              = watchdog,
 783         .thread_comm            = "watchdog/%u",
 784         .setup                  = watchdog_enable,
 785         .cleanup                = watchdog_cleanup,
 786         .park                   = watchdog_disable,
 787         .unpark                 = watchdog_enable,
 788 };
 789
 790 /*
 791  * park all watchdog threads that are specified in 'watchdog_cpumask'
 792  *
 793  * This function returns an error if kthread_park() of a watchdog thread
 794  * fails. In this situation, the watchdog threads of some CPUs can already
 795  * be parked and the watchdog threads of other CPUs can still be runnable.
 796  * Callers are expected to handle this special condition as appropriate in
 797  * their context.
 798  *
 799  * This function may only be called in a context that is protected against
 800  * races with CPU hotplug - for example, via get_online_cpus().
 801  */
 802 static int watchdog_park_threads(void)
 803 {
 804         int cpu, ret = 0;
 805
 806         for_each_watchdog_cpu(cpu) {
 807                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 808                 if (ret)
 809                         break;
 810         }
 811
 812         return ret;
 813 }
 814
 815 /*
 816  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 817  *
 818  * This function may only be called in a context that is protected against
 819  * races with CPU hotplug - for example, via get_online_cpus().
 820  */
 821 static void watchdog_unpark_threads(void)
 822 {
 823         int cpu;
 824
 825         for_each_watchdog_cpu(cpu)
 826                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 827 }
 828
 829 /*
 830  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 831  */
 832 int lockup_detector_suspend(void)
 833 {
 834         int ret = 0;
 835
 836         get_online_cpus();
 837         mutex_lock(&watchdog_proc_mutex);
 838         /*
 839          * Multiple suspend requests can be active in parallel (counted by
 840          * the 'watchdog_suspended' variable). If the watchdog threads are
 841          * running, the first caller takes care that they will be parked.
 842          * The state of 'watchdog_running' cannot change while a suspend
 843          * request is active (see related code in 'proc' handlers).
 844          */
 845         if (watchdog_running && !watchdog_suspended)
 846                 ret = watchdog_park_threads();
 847
 848         if (ret == 0)
 849                 watchdog_suspended++;
 850         else {
 851                 watchdog_disable_all_cpus();
 852                 pr_err("Failed to suspend lockup detectors, disabled\n");
 853                 watchdog_enabled = 0;
 854         }
 855
 856         mutex_unlock(&watchdog_proc_mutex);
 857
 858         return ret;
 859 }
 860
 861 /*
 862  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 863  */
 864 void lockup_detector_resume(void)
 865 {
 866         mutex_lock(&watchdog_proc_mutex);
 867
 868         watchdog_suspended--;
 869         /*
 870          * The watchdog threads are unparked if they were previously running
 871          * and if there is no more active suspend request.
 872          */
 873         if (watchdog_running && !watchdog_suspended)
 874                 watchdog_unpark_threads();
 875
 876         mutex_unlock(&watchdog_proc_mutex);
 877         put_online_cpus();
 878 }
 879
 880 static int update_watchdog_all_cpus(void)
 881 {
 882         int ret;
 883
 884         ret = watchdog_park_threads();
 885         if (ret)
 886                 return ret;
 887
 888         watchdog_unpark_threads();
 889
 890         return 0;
 891 }
 892
 893 static int watchdog_enable_all_cpus(void)
 894 {
 895         int err = 0;
 896
 897         if (!watchdog_running) {
 898                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 899                                                              &watchdog_cpumask);
 900                 if (err)
 901                         pr_err("Failed to create watchdog threads, disabled\n");
 902                 else
 903                         watchdog_running = 1;
 904         } else {
 905                 /*
 906                  * Enable/disable the lockup detectors or
 907                  * change the sample period 'on the fly'.
 908                  */
 909                 err = update_watchdog_all_cpus();
 910
 911                 if (err) {
 912                         watchdog_disable_all_cpus();
 913                         pr_err("Failed to update lockup detectors, disabled\n");
 914                 }
 915         }
 916
 917         if (err)
 918                 watchdog_enabled = 0;
 919
 920         return err;
 921 }
 922
 923 static void watchdog_disable_all_cpus(void)
 924 {
 925         if (watchdog_running) {
 926                 watchdog_running = 0;
 927                 smpboot_unregister_percpu_thread(&watchdog_threads);
 928         }
 929 }
 930
 931 #ifdef CONFIG_SYSCTL
 932
 933 /*
 934  * Update the run state of the lockup detectors.
 935  */
 936 static int proc_watchdog_update(void)
 937 {
 938         int err = 0;
 939
 940         /*
 941          * Watchdog threads won't be started if they are already active.
 942          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 943          * care of this. If those threads are already active, the sample
 944          * period will be updated and the lockup detectors will be enabled
 945          * or disabled 'on the fly'.
 946          */
 947         if (watchdog_enabled && watchdog_thresh)
 948                 err = watchdog_enable_all_cpus();
 949         else
 950                 watchdog_disable_all_cpus();
 951
 952         return err;
 953
 954 }
 955
 956 /*
 957  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 958  *
 959  * caller             | table->data points to | 'which' contains the flag(s)
 960  * -------------------|-----------------------|-----------------------------
 961  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 962  *                    |                       | with SOFT_WATCHDOG_ENABLED
 963  * -------------------|-----------------------|-----------------------------
 964  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 965  * -------------------|-----------------------|-----------------------------
 966  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 967  */
 968 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 969                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 970 {
 971         int err, old, new;
 972         int *watchdog_param = (int *)table->data;
 973
 974         get_online_cpus();
 975         mutex_lock(&watchdog_proc_mutex);
 976
 977         if (watchdog_suspended) {
 978                 /* no parameter changes allowed while watchdog is suspended */
 979                 err = -EAGAIN;
 980                 goto out;
 981         }
 982
 983         /*
 984          * If the parameter is being read return the state of the corresponding
 985          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 986          * run state of the lockup detectors.
 987          */
 988         if (!write) {
 989                 *watchdog_param = (watchdog_enabled & which) != 0;
 990                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 991         } else {
 992                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 993                 if (err)
 994                         goto out;
 995
 996                 /*
 997                  * There is a race window between fetching the current value
 998                  * from 'watchdog_enabled' and storing the new value. During
 999                  * this race window, watchdog_nmi_enable() can sneak in and
1000                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
1001                  * The 'cmpxchg' detects this race and the loop retries.
1002                  */
1003                 do {
1004                         old = watchdog_enabled;
1005                         /*
1006                          * If the parameter value is not zero set the
1007                          * corresponding bit(s), else clear it(them).
1008                          */
1009                         if (*watchdog_param)
1010                                 new = old | which;
1011                         else
1012                                 new = old & ~which;
1013                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
1014
1015                 /*
1016                  * Update the run state of the lockup detectors. There is _no_
1017                  * need to check the value returned by proc_watchdog_update()
1018                  * and to restore the previous value of 'watchdog_enabled' as
1019                  * both lockup detectors are disabled if proc_watchdog_update()
1020                  * returns an error.
1021                  */
1022                 if (old == new)
1023                         goto out;
1024
1025                 err = proc_watchdog_update();
1026         }
1027 out:
1028         mutex_unlock(&watchdog_proc_mutex);
1029         put_online_cpus();
1030         return err;
1031 }
1032
1033 /*
1034  * /proc/sys/kernel/watchdog
1035  */
1036 int proc_watchdog(struct ctl_table *table, int write,
1037                   void __user *buffer, size_t *lenp, loff_t *ppos)
1038 {
1039         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
1040                                     table, write, buffer, lenp, ppos);
1041 }
1042
1043 /*
1044  * /proc/sys/kernel/nmi_watchdog
1045  */
1046 int proc_nmi_watchdog(struct ctl_table *table, int write,
1047                       void __user *buffer, size_t *lenp, loff_t *ppos)
1048 {
1049         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
1050                                     table, write, buffer, lenp, ppos);
1051 }
1052
1053 /*
1054  * /proc/sys/kernel/soft_watchdog
1055  */
1056 int proc_soft_watchdog(struct ctl_table *table, int write,
1057                         void __user *buffer, size_t *lenp, loff_t *ppos)
1058 {
1059         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
1060                                     table, write, buffer, lenp, ppos);
1061 }
1062
1063 /*
1064  * /proc/sys/kernel/watchdog_thresh
1065  */
1066 int proc_watchdog_thresh(struct ctl_table *table, int write,
1067                          void __user *buffer, size_t *lenp, loff_t *ppos)
1068 {
1069         int err, old, new;
1070
1071         get_online_cpus();
1072         mutex_lock(&watchdog_proc_mutex);
1073
1074         if (watchdog_suspended) {
1075                 /* no parameter changes allowed while watchdog is suspended */
1076                 err = -EAGAIN;
1077                 goto out;
1078         }
1079
1080         old = ACCESS_ONCE(watchdog_thresh);
1081         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1082
1083         if (err || !write)
1084                 goto out;
1085
1086         /*
1087          * Update the sample period. Restore on failure.
1088          */
1089         new = ACCESS_ONCE(watchdog_thresh);
1090         if (old == new)
1091                 goto out;
1092
1093         set_sample_period();
1094         err = proc_watchdog_update();
1095         if (err) {
1096                 watchdog_thresh = old;
1097                 set_sample_period();
1098         }
1099 out:
1100         mutex_unlock(&watchdog_proc_mutex);
1101         put_online_cpus();
1102         return err;
1103 }
1104
1105 /*
1106  * The cpumask is the mask of possible cpus that the watchdog can run
1107  * on, not the mask of cpus it is actually running on.  This allows the
1108  * user to specify a mask that will include cpus that have not yet
1109  * been brought online, if desired.
1110  */
1111 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1112                           void __user *buffer, size_t *lenp, loff_t *ppos)
1113 {
1114         int err;
1115
1116         get_online_cpus();
1117         mutex_lock(&watchdog_proc_mutex);
1118
1119         if (watchdog_suspended) {
1120                 /* no parameter changes allowed while watchdog is suspended */
1121                 err = -EAGAIN;
1122                 goto out;
1123         }
1124
1125         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1126         if (!err && write) {
1127                 /* Remove impossible cpus to keep sysctl output cleaner. */
1128                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1129                             cpu_possible_mask);
1130
1131                 if (watchdog_running) {
1132                         /*
1133                          * Failure would be due to being unable to allocate
1134                          * a temporary cpumask, so we are likely not in a
1135                          * position to do much else to make things better.
1136                          */
1137                         if (smpboot_update_cpumask_percpu_thread(
1138                                     &watchdog_threads, &watchdog_cpumask) != 0)
1139                                 pr_err("cpumask update failed\n");
1140                 }
1141         }
1142 out:
1143         mutex_unlock(&watchdog_proc_mutex);
1144         put_online_cpus();
1145         return err;
1146 }
1147
1148 #endif /* CONFIG_SYSCTL */
1149
1150 void __init lockup_detector_init(void)
1151 {
1152         set_sample_period();
1153
1154 #ifdef CONFIG_NO_HZ_FULL
1155         if (tick_nohz_full_enabled()) {
1156                 pr_info("Disabling watchdog on nohz_full cores by default\n");
1157                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1158         } else
1159                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1160 #else
1161         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1162 #endif
1163
1164         if (watchdog_enabled)
1165                 watchdog_enable_all_cpus();
1166 }