kernel/watchdog_hld.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Detect hard lockups on a system
   4  *
   5  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   6  *
   7  * Note: Most of this code is borrowed heavily from the original softlockup
   8  * detector, so thanks to Ingo for the initial implementation.
   9  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
  10  * to those contributors as well.
  11  */
  12
  13 #define pr_fmt(fmt) "NMI watchdog: " fmt
  14
  15 #include <linux/nmi.h>
  16 #include <linux/atomic.h>
  17 #include <linux/module.h>
  18 #include <linux/sched/debug.h>
  19
  20 #include <asm/irq_regs.h>
  21 #include <linux/perf_event.h>
  22
  23 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  24 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  25 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  26 static DEFINE_PER_CPU(struct perf_event *, dead_event);
  27 static struct cpumask dead_events_mask;
  28
  29 static unsigned long hardlockup_allcpu_dumped;
  30 static atomic_t watchdog_cpus = ATOMIC_INIT(0);
  31
  32 notrace void arch_touch_nmi_watchdog(void)
  33 {
  34         /*
  35          * Using __raw here because some code paths have
  36          * preemption enabled.  If preemption is enabled
  37          * then interrupts should be enabled too, in which
  38          * case we shouldn't have to worry about the watchdog
  39          * going off.
  40          */
  41         raw_cpu_write(watchdog_nmi_touch, true);
  42 }
  43 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
  44
  45 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
  46 static DEFINE_PER_CPU(ktime_t, last_timestamp);
  47 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
  48 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
  49
  50 void watchdog_update_hrtimer_threshold(u64 period)
  51 {
  52         /*
  53          * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
  54          *
  55          * So it runs effectively with 2.5 times the rate of the NMI
  56          * watchdog. That means the hrtimer should fire 2-3 times before
  57          * the NMI watchdog expires. The NMI watchdog on x86 is based on
  58          * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
  59          * might run way faster than expected and the NMI fires in a
  60          * smaller period than the one deduced from the nominal CPU
  61          * frequency. Depending on the Turbo-Mode factor this might be fast
  62          * enough to get the NMI period smaller than the hrtimer watchdog
  63          * period and trigger false positives.
  64          *
  65          * The sample threshold is used to check in the NMI handler whether
  66          * the minimum time between two NMI samples has elapsed. That
  67          * prevents false positives.
  68          *
  69          * Set this to 4/5 of the actual watchdog threshold period so the
  70          * hrtimer is guaranteed to fire at least once within the real
  71          * watchdog threshold.
  72          */
  73         watchdog_hrtimer_sample_threshold = period * 2;
  74 }
  75
  76 static bool watchdog_check_timestamp(void)
  77 {
  78         ktime_t delta, now = ktime_get_mono_fast_ns();
  79
  80         delta = now - __this_cpu_read(last_timestamp);
  81         if (delta < watchdog_hrtimer_sample_threshold) {
  82                 /*
  83                  * If ktime is jiffies based, a stalled timer would prevent
  84                  * jiffies from being incremented and the filter would look
  85                  * at a stale timestamp and never trigger.
  86                  */
  87                 if (__this_cpu_inc_return(nmi_rearmed) < 10)
  88                         return false;
  89         }
  90         __this_cpu_write(nmi_rearmed, 0);
  91         __this_cpu_write(last_timestamp, now);
  92         return true;
  93 }
  94 #else
  95 static inline bool watchdog_check_timestamp(void)
  96 {
  97         return true;
  98 }
  99 #endif
 100
 101 static struct perf_event_attr wd_hw_attr = {
 102         .type           = PERF_TYPE_HARDWARE,
 103         .config         = PERF_COUNT_HW_CPU_CYCLES,
 104         .size           = sizeof(struct perf_event_attr),
 105         .pinned         = 1,
 106         .disabled       = 1,
 107 };
 108
 109 /* Callback function for perf event subsystem */
 110 static void watchdog_overflow_callback(struct perf_event *event,
 111                                        struct perf_sample_data *data,
 112                                        struct pt_regs *regs)
 113 {
 114         /* Ensure the watchdog never gets throttled */
 115         event->hw.interrupts = 0;
 116
 117         if (__this_cpu_read(watchdog_nmi_touch) == true) {
 118                 __this_cpu_write(watchdog_nmi_touch, false);
 119                 return;
 120         }
 121
 122         if (!watchdog_check_timestamp())
 123                 return;
 124
 125         /* check for a hardlockup
 126          * This is done by making sure our timer interrupt
 127          * is incrementing.  The timer interrupt should have
 128          * fired multiple times before we overflow'd.  If it hasn't
 129          * then this is a good indication the cpu is stuck
 130          */
 131         if (is_hardlockup()) {
 132                 int this_cpu = smp_processor_id();
 133
 134                 /* only print hardlockups once */
 135                 if (__this_cpu_read(hard_watchdog_warn) == true)
 136                         return;
 137
 138                 pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n",
 139                          this_cpu);
 140                 print_modules();
 141                 print_irqtrace_events(current);
 142                 if (regs)
 143                         show_regs(regs);
 144                 else
 145                         dump_stack();
 146
 147                 /*
 148                  * Perform all-CPU dump only once to avoid multiple hardlockups
 149                  * generating interleaving traces
 150                  */
 151                 if (sysctl_hardlockup_all_cpu_backtrace &&
 152                                 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
 153                         trigger_allbutself_cpu_backtrace();
 154
 155                 if (hardlockup_panic)
 156                         nmi_panic(regs, "Hard LOCKUP");
 157
 158                 __this_cpu_write(hard_watchdog_warn, true);
 159                 return;
 160         }
 161
 162         __this_cpu_write(hard_watchdog_warn, false);
 163         return;
 164 }
 165
 166 static int hardlockup_detector_event_create(void)
 167 {
 168         unsigned int cpu = smp_processor_id();
 169         struct perf_event_attr *wd_attr;
 170         struct perf_event *evt;
 171
 172         wd_attr = &wd_hw_attr;
 173         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 174
 175         /* Try to register using hardware perf events */
 176         evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
 177                                                watchdog_overflow_callback, NULL);
 178         if (IS_ERR(evt)) {
 179                 pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
 180                          PTR_ERR(evt));
 181                 return PTR_ERR(evt);
 182         }
 183         this_cpu_write(watchdog_ev, evt);
 184         return 0;
 185 }
 186
 187 /**
 188  * hardlockup_detector_perf_enable - Enable the local event
 189  */
 190 void hardlockup_detector_perf_enable(void)
 191 {
 192         if (hardlockup_detector_event_create())
 193                 return;
 194
 195         /* use original value for check */
 196         if (!atomic_fetch_inc(&watchdog_cpus))
 197                 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 198
 199         perf_event_enable(this_cpu_read(watchdog_ev));
 200 }
 201
 202 /**
 203  * hardlockup_detector_perf_disable - Disable the local event
 204  */
 205 void hardlockup_detector_perf_disable(void)
 206 {
 207         struct perf_event *event = this_cpu_read(watchdog_ev);
 208
 209         if (event) {
 210                 perf_event_disable(event);
 211                 this_cpu_write(watchdog_ev, NULL);
 212                 this_cpu_write(dead_event, event);
 213                 cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
 214                 atomic_dec(&watchdog_cpus);
 215         }
 216 }
 217
 218 /**
 219  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
 220  *
 221  * Called from lockup_detector_cleanup(). Serialized by the caller.
 222  */
 223 void hardlockup_detector_perf_cleanup(void)
 224 {
 225         int cpu;
 226
 227         for_each_cpu(cpu, &dead_events_mask) {
 228                 struct perf_event *event = per_cpu(dead_event, cpu);
 229
 230                 /*
 231                  * Required because for_each_cpu() reports  unconditionally
 232                  * CPU0 as set on UP kernels. Sigh.
 233                  */
 234                 if (event)
 235                         perf_event_release_kernel(event);
 236                 per_cpu(dead_event, cpu) = NULL;
 237         }
 238         cpumask_clear(&dead_events_mask);
 239 }
 240
 241 /**
 242  * hardlockup_detector_perf_stop - Globally stop watchdog events
 243  *
 244  * Special interface for x86 to handle the perf HT bug.
 245  */
 246 void __init hardlockup_detector_perf_stop(void)
 247 {
 248         int cpu;
 249
 250         lockdep_assert_cpus_held();
 251
 252         for_each_online_cpu(cpu) {
 253                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 254
 255                 if (event)
 256                         perf_event_disable(event);
 257         }
 258 }
 259
 260 /**
 261  * hardlockup_detector_perf_restart - Globally restart watchdog events
 262  *
 263  * Special interface for x86 to handle the perf HT bug.
 264  */
 265 void __init hardlockup_detector_perf_restart(void)
 266 {
 267         int cpu;
 268
 269         lockdep_assert_cpus_held();
 270
 271         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 272                 return;
 273
 274         for_each_online_cpu(cpu) {
 275                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 276
 277                 if (event)
 278                         perf_event_enable(event);
 279         }
 280 }
 281
 282 /**
 283  * hardlockup_detector_perf_init - Probe whether NMI event is available at all
 284  */
 285 int __init hardlockup_detector_perf_init(void)
 286 {
 287         int ret = hardlockup_detector_event_create();
 288
 289         if (ret) {
 290                 pr_info("Perf NMI watchdog permanently disabled\n");
 291         } else {
 292                 perf_event_release_kernel(this_cpu_read(watchdog_ev));
 293                 this_cpu_write(watchdog_ev, NULL);
 294         }
 295         return ret;
 296 }