kernel/watchdog_hld.c

   1 /*
   2  * Detect hard 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/nmi.h>
  15 #include <linux/module.h>
  16 #include <linux/sched/debug.h>
  17
  18 #include <asm/irq_regs.h>
  19 #include <linux/perf_event.h>
  20
  21 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  22 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  23 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  24 static DEFINE_PER_CPU(struct perf_event *, dead_event);
  25 static struct cpumask dead_events_mask;
  26
  27 static unsigned long hardlockup_allcpu_dumped;
  28 static bool hardlockup_detector_disabled;
  29
  30 void arch_touch_nmi_watchdog(void)
  31 {
  32         /*
  33          * Using __raw here because some code paths have
  34          * preemption enabled.  If preemption is enabled
  35          * then interrupts should be enabled too, in which
  36          * case we shouldn't have to worry about the watchdog
  37          * going off.
  38          */
  39         raw_cpu_write(watchdog_nmi_touch, true);
  40 }
  41 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
  42
  43 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
  44 static DEFINE_PER_CPU(ktime_t, last_timestamp);
  45 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
  46 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
  47
  48 void watchdog_update_hrtimer_threshold(u64 period)
  49 {
  50         /*
  51          * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
  52          *
  53          * So it runs effectively with 2.5 times the rate of the NMI
  54          * watchdog. That means the hrtimer should fire 2-3 times before
  55          * the NMI watchdog expires. The NMI watchdog on x86 is based on
  56          * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
  57          * might run way faster than expected and the NMI fires in a
  58          * smaller period than the one deduced from the nominal CPU
  59          * frequency. Depending on the Turbo-Mode factor this might be fast
  60          * enough to get the NMI period smaller than the hrtimer watchdog
  61          * period and trigger false positives.
  62          *
  63          * The sample threshold is used to check in the NMI handler whether
  64          * the minimum time between two NMI samples has elapsed. That
  65          * prevents false positives.
  66          *
  67          * Set this to 4/5 of the actual watchdog threshold period so the
  68          * hrtimer is guaranteed to fire at least once within the real
  69          * watchdog threshold.
  70          */
  71         watchdog_hrtimer_sample_threshold = period * 2;
  72 }
  73
  74 static bool watchdog_check_timestamp(void)
  75 {
  76         ktime_t delta, now = ktime_get_mono_fast_ns();
  77
  78         delta = now - __this_cpu_read(last_timestamp);
  79         if (delta < watchdog_hrtimer_sample_threshold) {
  80                 /*
  81                  * If ktime is jiffies based, a stalled timer would prevent
  82                  * jiffies from being incremented and the filter would look
  83                  * at a stale timestamp and never trigger.
  84                  */
  85                 if (__this_cpu_inc_return(nmi_rearmed) < 10)
  86                         return false;
  87         }
  88         __this_cpu_write(nmi_rearmed, 0);
  89         __this_cpu_write(last_timestamp, now);
  90         return true;
  91 }
  92 #else
  93 static inline bool watchdog_check_timestamp(void)
  94 {
  95         return true;
  96 }
  97 #endif
  98
  99 static struct perf_event_attr wd_hw_attr = {
 100         .type           = PERF_TYPE_HARDWARE,
 101         .config         = PERF_COUNT_HW_CPU_CYCLES,
 102         .size           = sizeof(struct perf_event_attr),
 103         .pinned         = 1,
 104         .disabled       = 1,
 105 };
 106
 107 /* Callback function for perf event subsystem */
 108 static void watchdog_overflow_callback(struct perf_event *event,
 109                                        struct perf_sample_data *data,
 110                                        struct pt_regs *regs)
 111 {
 112         /* Ensure the watchdog never gets throttled */
 113         event->hw.interrupts = 0;
 114
 115         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 116                 __this_cpu_write(watchdog_nmi_touch, false);
 117                 return;
 118         }
 119
 120         if (!watchdog_check_timestamp())
 121                 return;
 122
 123         /* check for a hardlockup
 124          * This is done by making sure our timer interrupt
 125          * is incrementing.  The timer interrupt should have
 126          * fired multiple times before we overflow'd.  If it hasn't
 127          * then this is a good indication the cpu is stuck
 128          */
 129         if (is_hardlockup()) {
 130                 int this_cpu = smp_processor_id();
 131
 132                 /* only print hardlockups once */
 133                 if (__this_cpu_read(hard_watchdog_warn) == true)
 134                         return;
 135
 136                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 137                 print_modules();
 138                 print_irqtrace_events(current);
 139                 if (regs)
 140                         show_regs(regs);
 141                 else
 142                         dump_stack();
 143
 144                 /*
 145                  * Perform all-CPU dump only once to avoid multiple hardlockups
 146                  * generating interleaving traces
 147                  */
 148                 if (sysctl_hardlockup_all_cpu_backtrace &&
 149                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 150                         trigger_allbutself_cpu_backtrace();
 151
 152                 if (hardlockup_panic)
 153                         nmi_panic(regs, "Hard LOCKUP");
 154
 155                 __this_cpu_write(hard_watchdog_warn, true);
 156                 return;
 157         }
 158
 159         __this_cpu_write(hard_watchdog_warn, false);
 160         return;
 161 }
 162
 163 /*
 164  * People like the simple clean cpu node info on boot.
 165  * Reduce the watchdog noise by only printing messages
 166  * that are different from what cpu0 displayed.
 167  */
 168 static unsigned long firstcpu_err;
 169 static atomic_t watchdog_cpus;
 170
 171 int watchdog_nmi_enable(unsigned int cpu)
 172 {
 173         struct perf_event_attr *wd_attr;
 174         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 175         int firstcpu = 0;
 176
 177         /* nothing to do if the hard lockup detector is disabled */
 178         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 179                 goto out;
 180
 181         /* A failure disabled the hardlockup detector permanently */
 182         if (hardlockup_detector_disabled)
 183                 return -ENODEV;
 184
 185         /* is it already setup and enabled? */
 186         if (event && event->state > PERF_EVENT_STATE_OFF)
 187                 goto out;
 188
 189         /* it is setup but not enabled */
 190         if (event != NULL)
 191                 goto out_enable;
 192
 193         if (atomic_inc_return(&watchdog_cpus) == 1)
 194                 firstcpu = 1;
 195
 196         wd_attr = &wd_hw_attr;
 197         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 198
 199         /* Try to register using hardware perf events */
 200         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 201
 202         /* save the first cpu's error for future comparision */
 203         if (firstcpu && IS_ERR(event))
 204                 firstcpu_err = PTR_ERR(event);
 205
 206         if (!IS_ERR(event)) {
 207                 /* only print for the first cpu initialized */
 208                 if (firstcpu || firstcpu_err)
 209                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 210                 goto out_save;
 211         }
 212
 213         /* skip displaying the same error again */
 214         if (!firstcpu && (PTR_ERR(event) == firstcpu_err))
 215                 return PTR_ERR(event);
 216
 217         /* vary the KERN level based on the returned errno */
 218         if (PTR_ERR(event) == -EOPNOTSUPP)
 219                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 220         else if (PTR_ERR(event) == -ENOENT)
 221                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 222                          cpu);
 223         else
 224                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 225                         cpu, PTR_ERR(event));
 226
 227         pr_info("Disabling hard lockup detector permanently\n");
 228         hardlockup_detector_disabled = true;
 229
 230         return PTR_ERR(event);
 231
 232         /* success path */
 233 out_save:
 234         per_cpu(watchdog_ev, cpu) = event;
 235 out_enable:
 236         perf_event_enable(per_cpu(watchdog_ev, cpu));
 237 out:
 238         return 0;
 239 }
 240
 241 static int hardlockup_detector_event_create(void)
 242 {
 243         unsigned int cpu = smp_processor_id();
 244         struct perf_event_attr *wd_attr;
 245         struct perf_event *evt;
 246
 247         wd_attr = &wd_hw_attr;
 248         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 249
 250         /* Try to register using hardware perf events */
 251         evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
 252                                                watchdog_overflow_callback, NULL);
 253         if (IS_ERR(evt)) {
 254                 pr_info("Perf event create on CPU %d failed with %ld\n", cpu,
 255                         PTR_ERR(evt));
 256                 return PTR_ERR(evt);
 257         }
 258         this_cpu_write(watchdog_ev, evt);
 259         return 0;
 260 }
 261
 262 /**
 263  * hardlockup_detector_perf_enable - Enable the local event
 264  */
 265 void hardlockup_detector_perf_enable(void)
 266 {
 267         if (hardlockup_detector_event_create())
 268                 return;
 269
 270         perf_event_enable(this_cpu_read(watchdog_ev));
 271 }
 272
 273 /**
 274  * hardlockup_detector_perf_disable - Disable the local event
 275  */
 276 void hardlockup_detector_perf_disable(void)
 277 {
 278         struct perf_event *event = this_cpu_read(watchdog_ev);
 279
 280         if (event) {
 281                 perf_event_disable(event);
 282                 this_cpu_write(watchdog_ev, NULL);
 283                 this_cpu_write(dead_event, event);
 284                 cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
 285
 286                 /* watchdog_nmi_enable() expects this to be zero initially. */
 287                 if (atomic_dec_and_test(&watchdog_cpus))
 288                         firstcpu_err = 0;
 289         }
 290 }
 291
 292 /**
 293  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
 294  *
 295  * Called from lockup_detector_cleanup(). Serialized by the caller.
 296  */
 297 void hardlockup_detector_perf_cleanup(void)
 298 {
 299         int cpu;
 300
 301         for_each_cpu(cpu, &dead_events_mask) {
 302                 struct perf_event *event = per_cpu(dead_event, cpu);
 303
 304                 per_cpu(dead_event, cpu) = NULL;
 305                 perf_event_release_kernel(event);
 306         }
 307         cpumask_clear(&dead_events_mask);
 308 }
 309
 310 /**
 311  * hardlockup_detector_perf_stop - Globally stop watchdog events
 312  *
 313  * Special interface for x86 to handle the perf HT bug.
 314  */
 315 void __init hardlockup_detector_perf_stop(void)
 316 {
 317         int cpu;
 318
 319         lockdep_assert_cpus_held();
 320
 321         for_each_online_cpu(cpu) {
 322                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 323
 324                 if (event)
 325                         perf_event_disable(event);
 326         }
 327 }
 328
 329 /**
 330  * hardlockup_detector_perf_restart - Globally restart watchdog events
 331  *
 332  * Special interface for x86 to handle the perf HT bug.
 333  */
 334 void __init hardlockup_detector_perf_restart(void)
 335 {
 336         int cpu;
 337
 338         lockdep_assert_cpus_held();
 339
 340         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 341                 return;
 342
 343         for_each_online_cpu(cpu) {
 344                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 345
 346                 if (event)
 347                         perf_event_enable(event);
 348         }
 349 }
 350
 351 /**
 352  * hardlockup_detector_perf_init - Probe whether NMI event is available at all
 353  */
 354 int __init hardlockup_detector_perf_init(void)
 355 {
 356         int ret = hardlockup_detector_event_create();
 357
 358         if (ret) {
 359                 pr_info("Perf NMI watchdog permanetely disabled\n");
 360         } else {
 361                 perf_event_release_kernel(this_cpu_read(watchdog_ev));
 362                 this_cpu_write(watchdog_ev, NULL);
 363         }
 364         return ret;
 365 }