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/delay.h>
  19 #include <linux/freezer.h>
  20 #include <linux/kthread.h>
  21 #include <linux/lockdep.h>
  22 #include <linux/notifier.h>
  23 #include <linux/module.h>
  24 #include <linux/sysctl.h>
  25
  26 #include <asm/irq_regs.h>
  27 #include <linux/kvm_para.h>
  28 #include <linux/perf_event.h>
  29
  30 int watchdog_enabled = 1;
  31 int __read_mostly watchdog_thresh = 10;
  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, soft_watchdog_warn);
  38 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  39 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  40 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  41 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  43 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  44 #endif
  45
  46 /* boot commands */
  47 /*
  48  * Should we panic when a soft-lockup or hard-lockup occurs:
  49  */
  50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  51 static int hardlockup_panic =
  52                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  53
  54 static int __init hardlockup_panic_setup(char *str)
  55 {
  56         if (!strncmp(str, "panic", 5))
  57                 hardlockup_panic = 1;
  58         else if (!strncmp(str, "nopanic", 7))
  59                 hardlockup_panic = 0;
  60         else if (!strncmp(str, "0", 1))
  61                 watchdog_enabled = 0;
  62         return 1;
  63 }
  64 __setup("nmi_watchdog=", hardlockup_panic_setup);
  65 #endif
  66
  67 unsigned int __read_mostly softlockup_panic =
  68                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  69
  70 static int __init softlockup_panic_setup(char *str)
  71 {
  72         softlockup_panic = simple_strtoul(str, NULL, 0);
  73
  74         return 1;
  75 }
  76 __setup("softlockup_panic=", softlockup_panic_setup);
  77
  78 static int __init nowatchdog_setup(char *str)
  79 {
  80         watchdog_enabled = 0;
  81         return 1;
  82 }
  83 __setup("nowatchdog", nowatchdog_setup);
  84
  85 /* deprecated */
  86 static int __init nosoftlockup_setup(char *str)
  87 {
  88         watchdog_enabled = 0;
  89         return 1;
  90 }
  91 __setup("nosoftlockup", nosoftlockup_setup);
  92 /*  */
  93
  94 /*
  95  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
  96  * lockups can have false positives under extreme conditions. So we generally
  97  * want a higher threshold for soft lockups than for hard lockups. So we couple
  98  * the thresholds with a factor: we make the soft threshold twice the amount of
  99  * time the hard threshold is.
 100  */
 101 static int get_softlockup_thresh(void)
 102 {
 103         return watchdog_thresh * 2;
 104 }
 105
 106 /*
 107  * Returns seconds, approximately.  We don't need nanosecond
 108  * resolution, and we don't need to waste time with a big divide when
 109  * 2^30ns == 1.074s.
 110  */
 111 static unsigned long get_timestamp(int this_cpu)
 112 {
 113         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 114 }
 115
 116 static unsigned long get_sample_period(void)
 117 {
 118         /*
 119          * convert watchdog_thresh from seconds to ns
 120          * the divide by 5 is to give hrtimer several chances (two
 121          * or three with the current relation between the soft
 122          * and hard thresholds) to increment before the
 123          * hardlockup detector generates 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
 285                 /* Clear the guest paused flag on watchdog reset */
 286                 kvm_check_and_clear_guest_paused();
 287                 __touch_watchdog();
 288                 return HRTIMER_RESTART;
 289         }
 290
 291         /* check for a softlockup
 292          * This is done by making sure a high priority task is
 293          * being scheduled.  The task touches the watchdog to
 294          * indicate it is getting cpu time.  If it hasn't then
 295          * this is a good indication some task is hogging the cpu
 296          */
 297         duration = is_softlockup(touch_ts);
 298         if (unlikely(duration)) {
 299                 /*
 300                  * If a virtual machine is stopped by the host it can look to
 301                  * the watchdog like a soft lockup, check to see if the host
 302                  * stopped the vm before we issue the warning
 303                  */
 304                 if (kvm_check_and_clear_guest_paused())
 305                         return HRTIMER_RESTART;
 306
 307                 /* only warn once */
 308                 if (__this_cpu_read(soft_watchdog_warn) == true)
 309                         return HRTIMER_RESTART;
 310
 311                 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 312                         smp_processor_id(), duration,
 313                         current->comm, task_pid_nr(current));
 314                 print_modules();
 315                 print_irqtrace_events(current);
 316                 if (regs)
 317                         show_regs(regs);
 318                 else
 319                         dump_stack();
 320
 321                 if (softlockup_panic)
 322                         panic("softlockup: hung tasks");
 323                 __this_cpu_write(soft_watchdog_warn, true);
 324         } else
 325                 __this_cpu_write(soft_watchdog_warn, false);
 326
 327         return HRTIMER_RESTART;
 328 }
 329
 330
 331 /*
 332  * The watchdog thread - touches the timestamp.
 333  */
 334 static int watchdog(void *unused)
 335 {
 336         struct sched_param param = { .sched_priority = 0 };
 337         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
 338
 339         /* initialize timestamp */
 340         __touch_watchdog();
 341
 342         /* kick off the timer for the hardlockup detector */
 343         /* done here because hrtimer_start can only pin to smp_processor_id() */
 344         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
 345                       HRTIMER_MODE_REL_PINNED);
 346
 347         set_current_state(TASK_INTERRUPTIBLE);
 348         /*
 349          * Run briefly (kicked by the hrtimer callback function) once every
 350          * get_sample_period() seconds (4 seconds by default) to reset the
 351          * softlockup timestamp. If this gets delayed for more than
 352          * 2*watchdog_thresh seconds then the debug-printout triggers in
 353          * watchdog_timer_fn().
 354          */
 355         while (!kthread_should_stop()) {
 356                 __touch_watchdog();
 357                 schedule();
 358
 359                 if (kthread_should_stop())
 360                         break;
 361
 362                 set_current_state(TASK_INTERRUPTIBLE);
 363         }
 364         /*
 365          * Drop the policy/priority elevation during thread exit to avoid a
 366          * scheduling latency spike.
 367          */
 368         __set_current_state(TASK_RUNNING);
 369         sched_setscheduler(current, SCHED_NORMAL, &param);
 370         return 0;
 371 }
 372
 373
 374 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 375 /*
 376  * People like the simple clean cpu node info on boot.
 377  * Reduce the watchdog noise by only printing messages
 378  * that are different from what cpu0 displayed.
 379  */
 380 static unsigned long cpu0_err;
 381
 382 static int watchdog_nmi_enable(int cpu)
 383 {
 384         struct perf_event_attr *wd_attr;
 385         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 386
 387         /* is it already setup and enabled? */
 388         if (event && event->state > PERF_EVENT_STATE_OFF)
 389                 goto out;
 390
 391         /* it is setup but not enabled */
 392         if (event != NULL)
 393                 goto out_enable;
 394
 395         wd_attr = &wd_hw_attr;
 396         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 397
 398         /* Try to register using hardware perf events */
 399         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 400
 401         /* save cpu0 error for future comparision */
 402         if (cpu == 0 && IS_ERR(event))
 403                 cpu0_err = PTR_ERR(event);
 404
 405         if (!IS_ERR(event)) {
 406                 /* only print for cpu0 or different than cpu0 */
 407                 if (cpu == 0 || cpu0_err)
 408                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 409                 goto out_save;
 410         }
 411
 412         /* skip displaying the same error again */
 413         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
 414                 return PTR_ERR(event);
 415
 416         /* vary the KERN level based on the returned errno */
 417         if (PTR_ERR(event) == -EOPNOTSUPP)
 418                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 419         else if (PTR_ERR(event) == -ENOENT)
 420                 pr_warning("disabled (cpu%i): hardware events not enabled\n",
 421                          cpu);
 422         else
 423                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 424                         cpu, PTR_ERR(event));
 425         return PTR_ERR(event);
 426
 427         /* success path */
 428 out_save:
 429         per_cpu(watchdog_ev, cpu) = event;
 430 out_enable:
 431         perf_event_enable(per_cpu(watchdog_ev, cpu));
 432 out:
 433         return 0;
 434 }
 435
 436 static void watchdog_nmi_disable(int cpu)
 437 {
 438         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 439
 440         if (event) {
 441                 perf_event_disable(event);
 442                 per_cpu(watchdog_ev, cpu) = NULL;
 443
 444                 /* should be in cleanup, but blocks oprofile */
 445                 perf_event_release_kernel(event);
 446         }
 447         return;
 448 }
 449 #else
 450 static int watchdog_nmi_enable(int cpu) { return 0; }
 451 static void watchdog_nmi_disable(int cpu) { return; }
 452 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
 453
 454 /* prepare/enable/disable routines */
 455 static void watchdog_prepare_cpu(int cpu)
 456 {
 457         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 458
 459         WARN_ON(per_cpu(softlockup_watchdog, cpu));
 460         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 461         hrtimer->function = watchdog_timer_fn;
 462 }
 463
 464 static int watchdog_enable(int cpu)
 465 {
 466         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 467         int err = 0;
 468
 469         /* enable the perf event */
 470         err = watchdog_nmi_enable(cpu);
 471
 472         /* Regardless of err above, fall through and start softlockup */
 473
 474         /* create the watchdog thread */
 475         if (!p) {
 476                 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 477                 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
 478                 if (IS_ERR(p)) {
 479                         pr_err("softlockup watchdog for %i failed\n", cpu);
 480                         if (!err) {
 481                                 /* if hardlockup hasn't already set this */
 482                                 err = PTR_ERR(p);
 483                                 /* and disable the perf event */
 484                                 watchdog_nmi_disable(cpu);
 485                         }
 486                         goto out;
 487                 }
 488                 sched_setscheduler(p, SCHED_FIFO, &param);
 489                 kthread_bind(p, cpu);
 490                 per_cpu(watchdog_touch_ts, cpu) = 0;
 491                 per_cpu(softlockup_watchdog, cpu) = p;
 492                 wake_up_process(p);
 493         }
 494
 495 out:
 496         return err;
 497 }
 498
 499 static void watchdog_disable(int cpu)
 500 {
 501         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
 502         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
 503
 504         /*
 505          * cancel the timer first to stop incrementing the stats
 506          * and waking up the kthread
 507          */
 508         hrtimer_cancel(hrtimer);
 509
 510         /* disable the perf event */
 511         watchdog_nmi_disable(cpu);
 512
 513         /* stop the watchdog thread */
 514         if (p) {
 515                 per_cpu(softlockup_watchdog, cpu) = NULL;
 516                 kthread_stop(p);
 517         }
 518 }
 519
 520 /* sysctl functions */
 521 #ifdef CONFIG_SYSCTL
 522 static void watchdog_enable_all_cpus(void)
 523 {
 524         int cpu;
 525
 526         watchdog_enabled = 0;
 527
 528         for_each_online_cpu(cpu)
 529                 if (!watchdog_enable(cpu))
 530                         /* if any cpu succeeds, watchdog is considered
 531                            enabled for the system */
 532                         watchdog_enabled = 1;
 533
 534         if (!watchdog_enabled)
 535                 pr_err("failed to be enabled on some cpus\n");
 536
 537 }
 538
 539 static void watchdog_disable_all_cpus(void)
 540 {
 541         int cpu;
 542
 543         for_each_online_cpu(cpu)
 544                 watchdog_disable(cpu);
 545
 546         /* if all watchdogs are disabled, then they are disabled for the system */
 547         watchdog_enabled = 0;
 548 }
 549
 550
 551 /*
 552  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
 553  */
 554
 555 int proc_dowatchdog(struct ctl_table *table, int write,
 556                     void __user *buffer, size_t *lenp, loff_t *ppos)
 557 {
 558         int ret;
 559
 560         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 561         if (ret || !write)
 562                 goto out;
 563
 564         if (watchdog_enabled && watchdog_thresh)
 565                 watchdog_enable_all_cpus();
 566         else
 567                 watchdog_disable_all_cpus();
 568
 569 out:
 570         return ret;
 571 }
 572 #endif /* CONFIG_SYSCTL */
 573
 574
 575 /*
 576  * Create/destroy watchdog threads as CPUs come and go:
 577  */
 578 static int
 579 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 580 {
 581         int hotcpu = (unsigned long)hcpu;
 582
 583         switch (action) {
 584         case CPU_UP_PREPARE:
 585         case CPU_UP_PREPARE_FROZEN:
 586                 watchdog_prepare_cpu(hotcpu);
 587                 break;
 588         case CPU_ONLINE:
 589         case CPU_ONLINE_FROZEN:
 590                 if (watchdog_enabled)
 591                         watchdog_enable(hotcpu);
 592                 break;
 593 #ifdef CONFIG_HOTPLUG_CPU
 594         case CPU_UP_CANCELED:
 595         case CPU_UP_CANCELED_FROZEN:
 596                 watchdog_disable(hotcpu);
 597                 break;
 598         case CPU_DEAD:
 599         case CPU_DEAD_FROZEN:
 600                 watchdog_disable(hotcpu);
 601                 break;
 602 #endif /* CONFIG_HOTPLUG_CPU */
 603         }
 604
 605         /*
 606          * hardlockup and softlockup are not important enough
 607          * to block cpu bring up.  Just always succeed and
 608          * rely on printk output to flag problems.
 609          */
 610         return NOTIFY_OK;
 611 }
 612
 613 static struct notifier_block cpu_nfb = {
 614         .notifier_call = cpu_callback
 615 };
 616
 617 #ifdef CONFIG_SUSPEND
 618 /*
 619  * On exit from suspend we force an offline->online transition on the boot CPU
 620  * so that the PMU state that was lost while in suspended state gets set up
 621  * properly for the boot CPU.  This information is required for restarting the
 622  * NMI watchdog.
 623  */
 624 void lockup_detector_bootcpu_resume(void)
 625 {
 626         void *cpu = (void *)(long)smp_processor_id();
 627
 628         cpu_callback(&cpu_nfb, CPU_DEAD_FROZEN, cpu);
 629         cpu_callback(&cpu_nfb, CPU_UP_PREPARE_FROZEN, cpu);
 630         cpu_callback(&cpu_nfb, CPU_ONLINE_FROZEN, cpu);
 631 }
 632 #endif
 633
 634 void __init lockup_detector_init(void)
 635 {
 636         void *cpu = (void *)(long)smp_processor_id();
 637         int err;
 638
 639         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 640         WARN_ON(notifier_to_errno(err));
 641
 642         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 643         register_cpu_notifier(&cpu_nfb);
 644
 645         return;
 646 }