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
  25 static unsigned long hardlockup_allcpu_dumped;
  26
  27 void arch_touch_nmi_watchdog(void)
  28 {
  29         /*
  30          * Using __raw here because some code paths have
  31          * preemption enabled.  If preemption is enabled
  32          * then interrupts should be enabled too, in which
  33          * case we shouldn't have to worry about the watchdog
  34          * going off.
  35          */
  36         raw_cpu_write(watchdog_nmi_touch, true);
  37 }
  38 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
  39
  40 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
  41 static DEFINE_PER_CPU(ktime_t, last_timestamp);
  42 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
  43 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
  44
  45 void watchdog_update_hrtimer_threshold(u64 period)
  46 {
  47         /*
  48          * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
  49          *
  50          * So it runs effectively with 2.5 times the rate of the NMI
  51          * watchdog. That means the hrtimer should fire 2-3 times before
  52          * the NMI watchdog expires. The NMI watchdog on x86 is based on
  53          * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
  54          * might run way faster than expected and the NMI fires in a
  55          * smaller period than the one deduced from the nominal CPU
  56          * frequency. Depending on the Turbo-Mode factor this might be fast
  57          * enough to get the NMI period smaller than the hrtimer watchdog
  58          * period and trigger false positives.
  59          *
  60          * The sample threshold is used to check in the NMI handler whether
  61          * the minimum time between two NMI samples has elapsed. That
  62          * prevents false positives.
  63          *
  64          * Set this to 4/5 of the actual watchdog threshold period so the
  65          * hrtimer is guaranteed to fire at least once within the real
  66          * watchdog threshold.
  67          */
  68         watchdog_hrtimer_sample_threshold = period * 2;
  69 }
  70
  71 static bool watchdog_check_timestamp(void)
  72 {
  73         ktime_t delta, now = ktime_get_mono_fast_ns();
  74
  75         delta = now - __this_cpu_read(last_timestamp);
  76         if (delta < watchdog_hrtimer_sample_threshold) {
  77                 /*
  78                  * If ktime is jiffies based, a stalled timer would prevent
  79                  * jiffies from being incremented and the filter would look
  80                  * at a stale timestamp and never trigger.
  81                  */
  82                 if (__this_cpu_inc_return(nmi_rearmed) < 10)
  83                         return false;
  84         }
  85         __this_cpu_write(nmi_rearmed, 0);
  86         __this_cpu_write(last_timestamp, now);
  87         return true;
  88 }
  89 #else
  90 static inline bool watchdog_check_timestamp(void)
  91 {
  92         return true;
  93 }
  94 #endif
  95
  96 static struct perf_event_attr wd_hw_attr = {
  97         .type           = PERF_TYPE_HARDWARE,
  98         .config         = PERF_COUNT_HW_CPU_CYCLES,
  99         .size           = sizeof(struct perf_event_attr),
 100         .pinned         = 1,
 101         .disabled       = 1,
 102 };
 103
 104 /* Callback function for perf event subsystem */
 105 static void watchdog_overflow_callback(struct perf_event *event,
 106                  struct perf_sample_data *data,
 107                  struct pt_regs *regs)
 108 {
 109         /* Ensure the watchdog never gets throttled */
 110         event->hw.interrupts = 0;
 111
 112         if (atomic_read(&watchdog_park_in_progress) != 0)
 113                 return;
 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         /* is it already setup and enabled? */
 182         if (event && event->state > PERF_EVENT_STATE_OFF)
 183                 goto out;
 184
 185         /* it is setup but not enabled */
 186         if (event != NULL)
 187                 goto out_enable;
 188
 189         if (atomic_inc_return(&watchdog_cpus) == 1)
 190                 firstcpu = 1;
 191
 192         wd_attr = &wd_hw_attr;
 193         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 194
 195         /* Try to register using hardware perf events */
 196         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
 197
 198         /* save the first cpu's error for future comparision */
 199         if (firstcpu && IS_ERR(event))
 200                 firstcpu_err = PTR_ERR(event);
 201
 202         if (!IS_ERR(event)) {
 203                 /* only print for the first cpu initialized */
 204                 if (firstcpu || firstcpu_err)
 205                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
 206                 goto out_save;
 207         }
 208
 209         /*
 210          * Disable the hard lockup detector if _any_ CPU fails to set up
 211          * set up the hardware perf event. The watchdog() function checks
 212          * the NMI_WATCHDOG_ENABLED bit periodically.
 213          *
 214          * The barriers are for syncing up watchdog_enabled across all the
 215          * cpus, as clear_bit() does not use barriers.
 216          */
 217         smp_mb__before_atomic();
 218         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
 219         smp_mb__after_atomic();
 220
 221         /* skip displaying the same error again */
 222         if (!firstcpu && (PTR_ERR(event) == firstcpu_err))
 223                 return PTR_ERR(event);
 224
 225         /* vary the KERN level based on the returned errno */
 226         if (PTR_ERR(event) == -EOPNOTSUPP)
 227                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
 228         else if (PTR_ERR(event) == -ENOENT)
 229                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
 230                          cpu);
 231         else
 232                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
 233                         cpu, PTR_ERR(event));
 234
 235         pr_info("Shutting down hard lockup detector on all cpus\n");
 236
 237         return PTR_ERR(event);
 238
 239         /* success path */
 240 out_save:
 241         per_cpu(watchdog_ev, cpu) = event;
 242 out_enable:
 243         perf_event_enable(per_cpu(watchdog_ev, cpu));
 244 out:
 245         return 0;
 246 }
 247
 248 void watchdog_nmi_disable(unsigned int cpu)
 249 {
 250         struct perf_event *event = per_cpu(watchdog_ev, cpu);
 251
 252         if (event) {
 253                 perf_event_disable(event);
 254                 per_cpu(watchdog_ev, cpu) = NULL;
 255
 256                 /* should be in cleanup, but blocks oprofile */
 257                 perf_event_release_kernel(event);
 258
 259                 /* watchdog_nmi_enable() expects this to be zero initially. */
 260                 if (atomic_dec_and_test(&watchdog_cpus))
 261                         firstcpu_err = 0;
 262         }
 263 }