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/module.h>
  23 #include <linux/lockdep.h>
  24 #include <linux/notifier.h>
  25 #include <linux/export.h>
  26 #include <linux/sysctl.h>
  27
  28 #include <asm/irq_regs.h>
  29 #include <linux/perf_event.h>
  30
  31 int watchdog_enabled = 1;
  32 int __read_mostly watchdog_thresh = 10;
  33
  34 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  35 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  36 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  37 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  38 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  39 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  40 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  41 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  43 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  44 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  45 #endif
  46
  47 /* boot commands */
  48 /*
  49  * Should we panic when a soft-lockup or hard-lockup occurs:
  50  */
  51 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  52 static int hardlockup_panic =
  53                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  54
  55 static int __init hardlockup_panic_setup(char *str)
  56 {
  57         if (!strncmp(str, "panic", 5))
  58                 hardlockup_panic = 1;
  59         else if (!strncmp(str, "nopanic", 7))
  60                 hardlockup_panic = 0;
  61         else if (!strncmp(str, "0", 1))
  62                 watchdog_enabled = 0;
  63         return 1;
  64 }
  65 __setup("nmi_watchdog=", hardlockup_panic_setup);
  66 #endif
  67
  68 unsigned int __read_mostly softlockup_panic =
  69                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  70
  71 static int __init softlockup_panic_setup(char *str)
  72 {
  73         softlockup_panic = simple_strtoul(str, NULL, 0);
  74
  75         return 1;
  76 }
  77 __setup("softlockup_panic=", softlockup_panic_setup);
  78
  79 static int __init nowatchdog_setup(char *str)
  80 {
  81         watchdog_enabled = 0;
  82         return 1;
  83 }
  84 __setup("nowatchdog", nowatchdog_setup);
  85
  86 /* deprecated */
  87 static int __init nosoftlockup_setup(char *str)
  88 {
  89         watchdog_enabled = 0;
  90         return 1;
  91 }
  92 __setup("nosoftlockup", nosoftlockup_setup);
  93 /*  */
  94
  95 /*
  96  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
  97  * lockups can have false positives under extreme conditions. So we generally
  98  * want a higher threshold for soft lockups than for hard lockups. So we couple
  99  * the thresholds with a factor: we make the soft threshold twice the amount of
 100  * time the hard threshold is.
 101  */
 102 static int get_softlockup_thresh(void)
 103 {
 104         return watchdog_thresh * 2;
 105 }
 106
 107 /*
 108  * Returns seconds, approximately.  We don't need nanosecond
 109  * resolution, and we don't need to waste time with a big divide when
 110  * 2^30ns == 1.074s.
 111  */
 112 static unsigned long get_timestamp(int this_cpu)
 113 {
 114         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 115 }
 116
 117 static unsigned long get_sample_period(void)
 118 {
 119         /*
 120          * convert watchdog_thresh from seconds to ns
 121          * the divide by 5 is to give hrtimer 5 chances to
 122          * increment before the hardlockup detector generates
 123          * a warning
 124          */
 125         return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
 126 }
 127
 128 /* Commands for resetting the watchdog */
 129 static void __touch_watchdog(void)
 130 {
 131         int this_cpu = smp_processor_id();
 132
 133         __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
 134 }
 135
 136 void touch_softlockup_watchdog(void)
 137 {
 138         __this_cpu_write(watchdog_touch_ts, 0);
 139 }
 140 EXPORT_SYMBOL(touch_softlockup_watchdog);
 141
 142 void touch_all_softlockup_watchdogs(void)
 143 {
 144         int cpu;
 145
 146         /*
 147          * this is done lockless
 148          * do we care if a 0 races with a timestamp?
 149          * all it means is the softlock check starts one cycle later
 150          */
 151         for_each_online_cpu(cpu)
 152                 per_cpu(watchdog_touch_ts, cpu) = 0;
 153 }
 154
 155 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 156 void touch_nmi_watchdog(void)
 157 {
 158         if (watchdog_enabled) {
 159                 unsigned cpu;
 160
 161                 for_each_present_cpu(cpu) {
 162                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
 163                                 per_cpu(watchdog_nmi_touch, cpu) = true;
 164                 }
 165         }
 166         touch_softlockup_watchdog();
 167 }
 168 EXPORT_SYMBOL(touch_nmi_watchdog);
 169
 170 #endif
 171
 172 void touch_softlockup_watchdog_sync(void)
 173 {
 174         __raw_get_cpu_var(softlockup_touch_sync) = true;
 175         __raw_get_cpu_var(watchdog_touch_ts) = 0;
 176 }
 177
 178 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 179 /* watchdog detector functions */
 180 static int is_hardlockup(void)
 181 {
 182         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 183
 184         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 185                 return 1;
 186
 187         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 188         return 0;
 189 }
 190 #endif
 191
 192 static int is_softlockup(unsigned long touch_ts)
 193 {
 194         unsigned long now = get_timestamp(smp_processor_id());
 195
 196         /* Warn about unreasonable delays: */
 197         if (time_after(now, touch_ts + get_softlockup_thresh()))
 198                 return now - touch_ts;
 199
 200         return 0;
 201 }
 202
 203 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 204
 205 static struct perf_event_attr wd_hw_attr = {
 206         .type           = PERF_TYPE_HARDWARE,
 207         .config         = PERF_COUNT_HW_CPU_CYCLES,
 208         .size           = sizeof(struct perf_event_attr),
 209         .pinned         = 1,
 210         .disabled       = 1,
 211 };
 212
 213 /* Callback function for perf event subsystem */
 214 static void watchdog_overflow_callback(struct perf_event *event,
 215                  struct perf_sample_data *data,
 216                  struct pt_regs *regs)
 217 {
 218         /* Ensure the watchdog never gets throttled */
 219         event->hw.interrupts = 0;
 220
 221         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 222                 __this_cpu_write(watchdog_nmi_touch, false);
 223                 return;
 224         }
 225
 226         /* check for a hardlockup
 227          * This is done by making sure our timer interrupt
 228          * is incrementing.  The timer interrupt should have
 229          * fired multiple times before we overflow'd.  If it hasn't
 230          * then this is a good indication the cpu is stuck
 231          */
 232         if (is_hardlockup()) {
 233                 int this_cpu = smp_processor_id();
 234
 235                 /* only print hardlockups once */
 236                 if (__this_cpu_read(hard_watchdog_warn) == true)
 237                         return;
 238
 239                 if (hardlockup_panic)
 240                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 241                 else
 242                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 243
 244                 __this_cpu_write(hard_watchdog_warn, true);
 245                 return;
 246         }
 247
 248         __this_cpu_write(hard_watchdog_warn, false);
 249         return;
 250 }
 251 static void watchdog_interrupt_count(void)
 252 {
 253         __this_cpu_inc(hrtimer_interrupts);
 254 }
 255 #else
 256 static inline void watchdog_interrupt_count(void) { return; }
 257 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 258
 259 /* watchdog kicker functions */
 260 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 261 {
 262         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 263         struct pt_regs *regs = get_irq_regs();
 264         int duration;
 265
 266         /* kick the hardlockup detector */
 267         watchdog_interrupt_count();
 268
 269         /* kick the softlockup detector */
 270         wake_up_process(__this_cpu_read(softlockup_watchdog));
 271
 272         /* .. and repeat */
 273         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
 274
 275         if (touch_ts == 0) {
 276                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 277                         /*
 278                          * If the time stamp was touched atomically
 279                          * make sure the scheduler tick is up to date.
 280                          */
 281                         __this_cpu_write(softlockup_touch_sync, false);
 282                         sched_clock_tick();
 283                 }
 284                 __touch_watchdog();
 285                 return HRTIMER_RESTART;
 286         }
 287
 288         /* check for a softlockup
 289          * This is done by making sure a high priority task is
 290          * being scheduled.  The task touches the watchdog to
 291          * indicate it is getting cpu time.  If it hasn't then
 292          * this is a good indication some task is hogging the cpu
 293          */
 294         duration = is_softlockup(touch_ts);
 295         if (unlikely(duration)) {
 296                 /* only warn once */
 297                 if (__this_cpu_read(soft_watchdog_warn) == true)
 298                         return HRTIMER_RESTART;
 299
 300                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 301                         smp_processor_id(), duration,
 302                         current->comm, task_pid_nr(current));
 303                 print_modules();
 304                 print_irqtrace_events(current);
 305                 if (regs)
 306                         show_regs(regs);
 307                 else
 308                         dump_stack();
 309
 310                 if (softlockup_panic)
 311                         panic("softlockup: hung tasks");
 312                 __this_cpu_write(soft_watchdog_warn, true);
 313         } else
 314                 __this_cpu_write(soft_watchdog_warn, false);
 315
 316         return HRTIMER_RESTART;
 317 }
 318
 319
 320 /*
 321  * The watchdog thread - touches the timestamp.
 322  */
 323 static int watchdog(void *unused)
 324 {
 325         static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 326         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 327
 328         sched_setscheduler(current, SCHED_FIFO, &param);
 329
 330         /* initialize timestamp */
 331         __touch_watchdog();
 332
 333         /* kick off the timer for the hardlockup detector */
 334         /* done here because hrtimer_start can only pin to smp_processor_id() */
 335         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
 336                       HRTIMER_MODE_REL_PINNED);
 337
 338         set_current_state(TASK_INTERRUPTIBLE);
 339         /*
 340          * Run briefly once per second to reset the softlockup timestamp.
 341          * If this gets delayed for more than 60 seconds then the
 342          * debug-printout triggers in watchdog_timer_fn().
 343          */
 344         while (!kthread_should_stop()) {
 345                 __touch_watchdog();
 346                 schedule();
 347
 348                 if (kthread_should_stop())
 349                         break;
 350
 351                 set_current_state(TASK_INTERRUPTIBLE);
 352         }
 353         __set_current_state(TASK_RUNNING);
 354
 355         return 0;
 356 }
 357
 358
 359 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 360 static int watchdog_nmi_enable(int cpu)
 361 {
 362         struct perf_event_attr *wd_attr;
 363         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 364
 365         /* is it already setup and enabled? */
 366         if (event && event->state > PERF_EVENT_STATE_OFF)
 367                 goto out;
 368
 369         /* it is setup but not enabled */
 370         if (event != NULL)
 371                 goto out_enable;
 372
 373         wd_attr = &wd_hw_attr;
 374         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 375
 376         /* Try to register using hardware perf events */
 377         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 378         if (!IS_ERR(event)) {
 379                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
 380                 goto out_save;
 381         }
 382
 383
 384         /* vary the KERN level based on the returned errno */
 385         if (PTR_ERR(event) == -EOPNOTSUPP)
 386                 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 387         else if (PTR_ERR(event) == -ENOENT)
 388                 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
 389         else
 390                 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
 391         return PTR_ERR(event);
 392
 393         /* success path */
 394 out_save:
 395         per_cpu(watchdog_ev, cpu) = event;
 396 out_enable:
 397         perf_event_enable(per_cpu(watchdog_ev, cpu));
 398 out:
 399         return 0;
 400 }
 401
 402 static void watchdog_nmi_disable(int cpu)
 403 {
 404         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 405
 406         if (event) {
 407                 perf_event_disable(event);
 408                 per_cpu(watchdog_ev, cpu) = NULL;
 409
 410                 /* should be in cleanup, but blocks oprofile */
 411                 perf_event_release_kernel(event);
 412         }
 413         return;
 414 }
 415 #else
 416 static int watchdog_nmi_enable(int cpu) { return 0; }
 417 static void watchdog_nmi_disable(int cpu) { return; }
 418 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 419
 420 /* prepare/enable/disable routines */
 421 static void watchdog_prepare_cpu(int cpu)
 422 {
 423         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 424
 425         WARN_ON(per_cpu(softlockup_watchdog, cpu));
 426         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 427         hrtimer->function = watchdog_timer_fn;
 428 }
 429
 430 static int watchdog_enable(int cpu)
 431 {
 432         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 433         int err = 0;
 434
 435         /* enable the perf event */
 436         err = watchdog_nmi_enable(cpu);
 437
 438         /* Regardless of err above, fall through and start softlockup */
 439
 440         /* create the watchdog thread */
 441         if (!p) {
 442                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
 443                 if (IS_ERR(p)) {
 444                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
 445                         if (!err) {
 446                                 /* if hardlockup hasn't already set this */
 447                                 err = PTR_ERR(p);
 448                                 /* and disable the perf event */
 449                                 watchdog_nmi_disable(cpu);
 450                         }
 451                         goto out;
 452                 }
 453                 kthread_bind(p, cpu);
 454                 per_cpu(watchdog_touch_ts, cpu) = 0;
 455                 per_cpu(softlockup_watchdog, cpu) = p;
 456                 wake_up_process(p);
 457         }
 458
 459 out:
 460         return err;
 461 }
 462
 463 static void watchdog_disable(int cpu)
 464 {
 465         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 466         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 467
 468         /*
 469          * cancel the timer first to stop incrementing the stats
 470          * and waking up the kthread
 471          */
 472         hrtimer_cancel(hrtimer);
 473
 474         /* disable the perf event */
 475         watchdog_nmi_disable(cpu);
 476
 477         /* stop the watchdog thread */
 478         if (p) {
 479                 per_cpu(softlockup_watchdog, cpu) = NULL;
 480                 kthread_stop(p);
 481         }
 482 }
 483
 484 static void watchdog_enable_all_cpus(void)
 485 {
 486         int cpu;
 487
 488         watchdog_enabled = 0;
 489
 490         for_each_online_cpu(cpu)
 491                 if (!watchdog_enable(cpu))
 492                         /* if any cpu succeeds, watchdog is considered
 493                            enabled for the system */
 494                         watchdog_enabled = 1;
 495
 496         if (!watchdog_enabled)
 497                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
 498
 499 }
 500
 501 static void watchdog_disable_all_cpus(void)
 502 {
 503         int cpu;
 504
 505         for_each_online_cpu(cpu)
 506                 watchdog_disable(cpu);
 507
 508         /* if all watchdogs are disabled, then they are disabled for the system */
 509         watchdog_enabled = 0;
 510 }
 511
 512
 513 /* sysctl functions */
 514 #ifdef CONFIG_SYSCTL
 515 /*
 516  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 517  */
 518
 519 int proc_dowatchdog(struct ctl_table *table, int write,
 520                     void __user *buffer, size_t *lenp, loff_t *ppos)
 521 {
 522         int ret;
 523
 524         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 525         if (ret || !write)
 526                 goto out;
 527
 528         if (watchdog_enabled && watchdog_thresh)
 529                 watchdog_enable_all_cpus();
 530         else
 531                 watchdog_disable_all_cpus();
 532
 533 out:
 534         return ret;
 535 }
 536 #endif /* CONFIG_SYSCTL */
 537
 538
 539 /*
 540  * Create/destroy watchdog threads as CPUs come and go:
 541  */
 542 static int __cpuinit
 543 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 544 {
 545         int hotcpu = (unsigned long)hcpu;
 546
 547         switch (action) {
 548         case CPU_UP_PREPARE:
 549         case CPU_UP_PREPARE_FROZEN:
 550                 watchdog_prepare_cpu(hotcpu);
 551                 break;
 552         case CPU_ONLINE:
 553         case CPU_ONLINE_FROZEN:
 554                 if (watchdog_enabled)
 555                         watchdog_enable(hotcpu);
 556                 break;
 557 #ifdef CONFIG_HOTPLUG_CPU
 558         case CPU_UP_CANCELED:
 559         case CPU_UP_CANCELED_FROZEN:
 560                 watchdog_disable(hotcpu);
 561                 break;
 562         case CPU_DEAD:
 563         case CPU_DEAD_FROZEN:
 564                 watchdog_disable(hotcpu);
 565                 break;
 566 #endif /* CONFIG_HOTPLUG_CPU */
 567         }
 568
 569         /*
 570          * hardlockup and softlockup are not important enough
 571          * to block cpu bring up.  Just always succeed and
 572          * rely on printk output to flag problems.
 573          */
 574         return NOTIFY_OK;
 575 }
 576
 577 static struct notifier_block __cpuinitdata cpu_nfb = {
 578         .notifier_call = cpu_callback
 579 };
 580
 581 void __init lockup_detector_init(void)
 582 {
 583         void *cpu = (void *)(long)smp_processor_id();
 584         int err;
 585
 586         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 587         WARN_ON(notifier_to_errno(err));
 588
 589         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 590         register_cpu_notifier(&cpu_nfb);
 591
 592         return;
 593 }