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  * this code detects hard lockups: incidents in where on a CPU
   7  * the kernel does not respond to anything except NMI.
   8  *
   9  * Note: Most of this code is borrowed heavily from softlockup.c,
  10  * so thanks to Ingo for the initial implementation.
  11  * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
  12  * to those contributors as well.
  13  */
  14
  15 #include <linux/mm.h>
  16 #include <linux/cpu.h>
  17 #include <linux/nmi.h>
  18 #include <linux/init.h>
  19 #include <linux/delay.h>
  20 #include <linux/freezer.h>
  21 #include <linux/kthread.h>
  22 #include <linux/lockdep.h>
  23 #include <linux/notifier.h>
  24 #include <linux/module.h>
  25 #include <linux/sysctl.h>
  26
  27 #include <asm/irq_regs.h>
  28 #include <linux/perf_event.h>
  29
  30 int watchdog_enabled;
  31 int __read_mostly softlockup_thresh = 60;
  32
  33 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  34 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  35 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  36 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  37 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  38 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  39 #ifdef CONFIG_PERF_EVENTS_NMI
  40 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  41 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  42 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  43 #endif
  44
  45 static int __read_mostly did_panic;
  46 static int __initdata no_watchdog;
  47
  48
  49 /* boot commands */
  50 /*
  51  * Should we panic when a soft-lockup or hard-lockup occurs:
  52  */
  53 #ifdef CONFIG_PERF_EVENTS_NMI
  54 static int hardlockup_panic;
  55
  56 static int __init hardlockup_panic_setup(char *str)
  57 {
  58         if (!strncmp(str, "panic", 5))
  59                 hardlockup_panic = 1;
  60         return 1;
  61 }
  62 __setup("nmi_watchdog=", hardlockup_panic_setup);
  63 #endif
  64
  65 unsigned int __read_mostly softlockup_panic =
  66                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  67
  68 static int __init softlockup_panic_setup(char *str)
  69 {
  70         softlockup_panic = simple_strtoul(str, NULL, 0);
  71
  72         return 1;
  73 }
  74 __setup("softlockup_panic=", softlockup_panic_setup);
  75
  76 static int __init nowatchdog_setup(char *str)
  77 {
  78         no_watchdog = 1;
  79         return 1;
  80 }
  81 __setup("nowatchdog", nowatchdog_setup);
  82
  83 /* deprecated */
  84 static int __init nosoftlockup_setup(char *str)
  85 {
  86         no_watchdog = 1;
  87         return 1;
  88 }
  89 __setup("nosoftlockup", nosoftlockup_setup);
  90 /*  */
  91
  92
  93 /*
  94  * Returns seconds, approximately.  We don't need nanosecond
  95  * resolution, and we don't need to waste time with a big divide when
  96  * 2^30ns == 1.074s.
  97  */
  98 static unsigned long get_timestamp(int this_cpu)
  99 {
 100         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 101 }
 102
 103 static unsigned long get_sample_period(void)
 104 {
 105         /*
 106          * convert softlockup_thresh from seconds to ns
 107          * the divide by 5 is to give hrtimer 5 chances to
 108          * increment before the hardlockup detector generates
 109          * a warning
 110          */
 111         return softlockup_thresh / 5 * NSEC_PER_SEC;
 112 }
 113
 114 /* Commands for resetting the watchdog */
 115 static void __touch_watchdog(void)
 116 {
 117         int this_cpu = raw_smp_processor_id();
 118
 119         __get_cpu_var(watchdog_touch_ts) = get_timestamp(this_cpu);
 120 }
 121
 122 void touch_watchdog(void)
 123 {
 124         __get_cpu_var(watchdog_touch_ts) = 0;
 125 }
 126 EXPORT_SYMBOL(touch_watchdog);
 127
 128 void touch_all_watchdog(void)
 129 {
 130         int cpu;
 131
 132         /*
 133          * this is done lockless
 134          * do we care if a 0 races with a timestamp?
 135          * all it means is the softlock check starts one cycle later
 136          */
 137         for_each_online_cpu(cpu)
 138                 per_cpu(watchdog_touch_ts, cpu) = 0;
 139 }
 140
 141 void touch_nmi_watchdog(void)
 142 {
 143         touch_watchdog();
 144 }
 145 EXPORT_SYMBOL(touch_nmi_watchdog);
 146
 147 void touch_all_nmi_watchdog(void)
 148 {
 149         touch_all_watchdog();
 150 }
 151
 152 void touch_softlockup_watchdog(void)
 153 {
 154         touch_watchdog();
 155 }
 156
 157 void touch_all_softlockup_watchdogs(void)
 158 {
 159         touch_all_watchdog();
 160 }
 161
 162 void touch_softlockup_watchdog_sync(void)
 163 {
 164         __raw_get_cpu_var(softlockup_touch_sync) = true;
 165         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 166 }
 167
 168 void softlockup_tick(void)
 169 {
 170 }
 171
 172 #ifdef CONFIG_PERF_EVENTS_NMI
 173 /* watchdog detector functions */
 174 static int is_hardlockup(int cpu)
 175 {
 176         unsigned long hrint = per_cpu(hrtimer_interrupts, cpu);
 177
 178         if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
 179                 return 1;
 180
 181         per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
 182         return 0;
 183 }
 184 #endif
 185
 186 static int is_softlockup(unsigned long touch_ts, int cpu)
 187 {
 188         unsigned long now = get_timestamp(cpu);
 189
 190         /* Warn about unreasonable delays: */
 191         if (time_after(now, touch_ts + softlockup_thresh))
 192                 return now - touch_ts;
 193
 194         return 0;
 195 }
 196
 197 static int
 198 watchdog_panic(struct notifier_block *this, unsigned long event, void *ptr)
 199 {
 200         did_panic = 1;
 201
 202         return NOTIFY_DONE;
 203 }
 204
 205 static struct notifier_block panic_block = {
 206         .notifier_call = watchdog_panic,
 207 };
 208
 209 #ifdef CONFIG_PERF_EVENTS_NMI
 210 static struct perf_event_attr wd_hw_attr = {
 211         .type           = PERF_TYPE_HARDWARE,
 212         .config         = PERF_COUNT_HW_CPU_CYCLES,
 213         .size           = sizeof(struct perf_event_attr),
 214         .pinned         = 1,
 215         .disabled       = 1,
 216 };
 217
 218 /* Callback function for perf event subsystem */
 219 void watchdog_overflow_callback(struct perf_event *event, int nmi,
 220                  struct perf_sample_data *data,
 221                  struct pt_regs *regs)
 222 {
 223         int this_cpu = smp_processor_id();
 224         unsigned long touch_ts = per_cpu(watchdog_touch_ts, this_cpu);
 225
 226         if (touch_ts == 0) {
 227                 __touch_watchdog();
 228                 return;
 229         }
 230
 231         /* check for a hardlockup
 232          * This is done by making sure our timer interrupt
 233          * is incrementing.  The timer interrupt should have
 234          * fired multiple times before we overflow'd.  If it hasn't
 235          * then this is a good indication the cpu is stuck
 236          */
 237         if (is_hardlockup(this_cpu)) {
 238                 /* only print hardlockups once */
 239                 if (__get_cpu_var(hard_watchdog_warn) == true)
 240                         return;
 241
 242                 if (hardlockup_panic)
 243                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 244                 else
 245                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 246
 247                 __get_cpu_var(hard_watchdog_warn) = true;
 248                 return;
 249         }
 250
 251         __get_cpu_var(hard_watchdog_warn) = false;
 252         return;
 253 }
 254 static void watchdog_interrupt_count(void)
 255 {
 256         __get_cpu_var(hrtimer_interrupts)++;
 257 }
 258 #else
 259 static inline void watchdog_interrupt_count(void) { return; }
 260 #endif /* CONFIG_PERF_EVENTS_NMI */
 261
 262 /* watchdog kicker functions */
 263 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 264 {
 265         int this_cpu = smp_processor_id();
 266         unsigned long touch_ts = __get_cpu_var(watchdog_touch_ts);
 267         struct pt_regs *regs = get_irq_regs();
 268         int duration;
 269
 270         /* kick the hardlockup detector */
 271         watchdog_interrupt_count();
 272
 273         /* kick the softlockup detector */
 274         wake_up_process(__get_cpu_var(softlockup_watchdog));
 275
 276         /* .. and repeat */
 277         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
 278
 279         if (touch_ts == 0) {
 280                 if (unlikely(per_cpu(softlockup_touch_sync, this_cpu))) {
 281                         /*
 282                          * If the time stamp was touched atomically
 283                          * make sure the scheduler tick is up to date.
 284                          */
 285                         per_cpu(softlockup_touch_sync, this_cpu) = false;
 286                         sched_clock_tick();
 287                 }
 288                 __touch_watchdog();
 289                 return HRTIMER_RESTART;
 290         }
 291
 292         /* check for a softlockup
 293          * This is done by making sure a high priority task is
 294          * being scheduled.  The task touches the watchdog to
 295          * indicate it is getting cpu time.  If it hasn't then
 296          * this is a good indication some task is hogging the cpu
 297          */
 298         duration = is_softlockup(touch_ts, this_cpu);
 299         if (unlikely(duration)) {
 300                 /* only warn once */
 301                 if (__get_cpu_var(soft_watchdog_warn) == true)
 302                         return HRTIMER_RESTART;
 303
 304                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 305                         this_cpu, duration,
 306                         current->comm, task_pid_nr(current));
 307                 print_modules();
 308                 print_irqtrace_events(current);
 309                 if (regs)
 310                         show_regs(regs);
 311                 else
 312                         dump_stack();
 313
 314                 if (softlockup_panic)
 315                         panic("softlockup: hung tasks");
 316                 __get_cpu_var(soft_watchdog_warn) = true;
 317         } else
 318                 __get_cpu_var(soft_watchdog_warn) = false;
 319
 320         return HRTIMER_RESTART;
 321 }
 322
 323
 324 /*
 325  * The watchdog thread - touches the timestamp.
 326  */
 327 static int watchdog(void *__bind_cpu)
 328 {
 329         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 330         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, (unsigned long)__bind_cpu);
 331
 332         sched_setscheduler(current, SCHED_FIFO, &param);
 333
 334         /* initialize timestamp */
 335         __touch_watchdog();
 336
 337         /* kick off the timer for the hardlockup detector */
 338         /* done here because hrtimer_start can only pin to smp_processor_id() */
 339         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
 340                       HRTIMER_MODE_REL_PINNED);
 341
 342         set_current_state(TASK_INTERRUPTIBLE);
 343         /*
 344          * Run briefly once per second to reset the softlockup timestamp.
 345          * If this gets delayed for more than 60 seconds then the
 346          * debug-printout triggers in softlockup_tick().
 347          */
 348         while (!kthread_should_stop()) {
 349                 __touch_watchdog();
 350                 schedule();
 351
 352                 if (kthread_should_stop())
 353                         break;
 354
 355                 set_current_state(TASK_INTERRUPTIBLE);
 356         }
 357         __set_current_state(TASK_RUNNING);
 358
 359         return 0;
 360 }
 361
 362
 363 #ifdef CONFIG_PERF_EVENTS_NMI
 364 static int watchdog_nmi_enable(int cpu)
 365 {
 366         struct perf_event_attr *wd_attr;
 367         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 368
 369         /* is it already setup and enabled? */
 370         if (event && event->state > PERF_EVENT_STATE_OFF)
 371                 goto out;
 372
 373         /* it is setup but not enabled */
 374         if (event != NULL)
 375                 goto out_enable;
 376
 377         /* Try to register using hardware perf events */
 378         wd_attr = &wd_hw_attr;
 379         wd_attr->sample_period = hw_nmi_get_sample_period();
 380         event = perf_event_create_kernel_counter(wd_attr, cpu, -1, watchdog_overflow_callback);
 381         if (!IS_ERR(event)) {
 382                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
 383                 goto out_save;
 384         }
 385
 386         printk(KERN_ERR "NMI watchdog failed to create perf event on cpu%i: %p\n", cpu, event);
 387         return -1;
 388
 389         /* success path */
 390 out_save:
 391         per_cpu(watchdog_ev, cpu) = event;
 392 out_enable:
 393         perf_event_enable(per_cpu(watchdog_ev, cpu));
 394 out:
 395         return 0;
 396 }
 397
 398 static void watchdog_nmi_disable(int cpu)
 399 {
 400         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 401
 402         if (event) {
 403                 perf_event_disable(event);
 404                 per_cpu(watchdog_ev, cpu) = NULL;
 405
 406                 /* should be in cleanup, but blocks oprofile */
 407                 perf_event_release_kernel(event);
 408         }
 409         return;
 410 }
 411 #else
 412 static int watchdog_nmi_enable(int cpu) { return 0; }
 413 static void watchdog_nmi_disable(int cpu) { return; }
 414 #endif /* CONFIG_PERF_EVENTS_NMI */
 415
 416 /* prepare/enable/disable routines */
 417 static int watchdog_prepare_cpu(int cpu)
 418 {
 419         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 420
 421         WARN_ON(per_cpu(softlockup_watchdog, cpu));
 422         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 423         hrtimer->function = watchdog_timer_fn;
 424
 425         return 0;
 426 }
 427
 428 static int watchdog_enable(int cpu)
 429 {
 430         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 431
 432         /* enable the perf event */
 433         if (watchdog_nmi_enable(cpu) != 0)
 434                 return -1;
 435
 436         /* create the watchdog thread */
 437         if (!p) {
 438                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
 439                 if (IS_ERR(p)) {
 440                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
 441                         return -1;
 442                 }
 443                 kthread_bind(p, cpu);
 444                 per_cpu(watchdog_touch_ts, cpu) = 0;
 445                 per_cpu(softlockup_watchdog, cpu) = p;
 446                 wake_up_process(p);
 447         }
 448
 449         return 0;
 450 }
 451
 452 static void watchdog_disable(int cpu)
 453 {
 454         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 455         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 456
 457         /*
 458          * cancel the timer first to stop incrementing the stats
 459          * and waking up the kthread
 460          */
 461         hrtimer_cancel(hrtimer);
 462
 463         /* disable the perf event */
 464         watchdog_nmi_disable(cpu);
 465
 466         /* stop the watchdog thread */
 467         if (p) {
 468                 per_cpu(softlockup_watchdog, cpu) = NULL;
 469                 kthread_stop(p);
 470         }
 471
 472         /* if any cpu succeeds, watchdog is considered enabled for the system */
 473         watchdog_enabled = 1;
 474 }
 475
 476 static void watchdog_enable_all_cpus(void)
 477 {
 478         int cpu;
 479         int result;
 480
 481         for_each_online_cpu(cpu)
 482                 result += watchdog_enable(cpu);
 483
 484         if (result)
 485                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
 486
 487 }
 488
 489 static void watchdog_disable_all_cpus(void)
 490 {
 491         int cpu;
 492
 493         for_each_online_cpu(cpu)
 494                 watchdog_disable(cpu);
 495
 496         /* if all watchdogs are disabled, then they are disabled for the system */
 497         watchdog_enabled = 0;
 498 }
 499
 500
 501 /* sysctl functions */
 502 #ifdef CONFIG_SYSCTL
 503 /*
 504  * proc handler for /proc/sys/kernel/nmi_watchdog
 505  */
 506
 507 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
 508                      void __user *buffer, size_t *length, loff_t *ppos)
 509 {
 510         proc_dointvec(table, write, buffer, length, ppos);
 511
 512         if (watchdog_enabled)
 513                 watchdog_enable_all_cpus();
 514         else
 515                 watchdog_disable_all_cpus();
 516         return 0;
 517 }
 518
 519 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
 520                              void __user *buffer,
 521                              size_t *lenp, loff_t *ppos)
 522 {
 523         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 524 }
 525
 526 /* stub functions */
 527 int proc_dosoftlockup_thresh(struct ctl_table *table, int write,
 528                              void __user *buffer,
 529                              size_t *lenp, loff_t *ppos)
 530 {
 531         return proc_dowatchdog_thresh(table, write, buffer, lenp, ppos);
 532 }
 533 /* end of stub functions */
 534 #endif /* CONFIG_SYSCTL */
 535
 536
 537 /*
 538  * Create/destroy watchdog threads as CPUs come and go:
 539  */
 540 static int __cpuinit
 541 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 542 {
 543         int hotcpu = (unsigned long)hcpu;
 544
 545         switch (action) {
 546         case CPU_UP_PREPARE:
 547         case CPU_UP_PREPARE_FROZEN:
 548                 if (watchdog_prepare_cpu(hotcpu))
 549                         return NOTIFY_BAD;
 550                 break;
 551         case CPU_ONLINE:
 552         case CPU_ONLINE_FROZEN:
 553                 if (watchdog_enable(hotcpu))
 554                         return NOTIFY_BAD;
 555                 break;
 556 #ifdef CONFIG_HOTPLUG_CPU
 557         case CPU_UP_CANCELED:
 558         case CPU_UP_CANCELED_FROZEN:
 559                 watchdog_disable(hotcpu);
 560                 break;
 561         case CPU_DEAD:
 562         case CPU_DEAD_FROZEN:
 563                 watchdog_disable(hotcpu);
 564                 break;
 565 #endif /* CONFIG_HOTPLUG_CPU */
 566         }
 567         return NOTIFY_OK;
 568 }
 569
 570 static struct notifier_block __cpuinitdata cpu_nfb = {
 571         .notifier_call = cpu_callback
 572 };
 573
 574 static int __init spawn_watchdog_task(void)
 575 {
 576         void *cpu = (void *)(long)smp_processor_id();
 577         int err;
 578
 579         if (no_watchdog)
 580                 return 0;
 581
 582         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 583         WARN_ON(err == NOTIFY_BAD);
 584
 585         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 586         register_cpu_notifier(&cpu_nfb);
 587
 588         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 589
 590         return 0;
 591 }
 592 early_initcall(spawn_watchdog_task);