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/module.h>
  17 #include <linux/sched/debug.h>
  18
  19 #include <asm/irq_regs.h>
  20 #include <linux/perf_event.h>
  21
  22 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  23 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  24 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  25 static struct cpumask dead_events_mask;
  26
  27 static unsigned long hardlockup_allcpu_dumped;
  28 static unsigned int watchdog_cpus;
  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 static int hardlockup_detector_event_create(void)
 164 {
 165         unsigned int cpu = smp_processor_id();
 166         struct perf_event_attr *wd_attr;
 167         struct perf_event *evt;
 168
 169         wd_attr = &wd_hw_attr;
 170         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 171
 172         /* Try to register using hardware perf events */
 173         evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
 174                                                watchdog_overflow_callback, NULL);
 175         if (IS_ERR(evt)) {
 176                 pr_info("Perf event create on CPU %d failed with %ld\n", cpu,
 177                         PTR_ERR(evt));
 178                 return PTR_ERR(evt);
 179         }
 180         this_cpu_write(watchdog_ev, evt);
 181         return 0;
 182 }
 183
 184 /**
 185  * hardlockup_detector_perf_enable - Enable the local event
 186  */
 187 void hardlockup_detector_perf_enable(void)
 188 {
 189         if (hardlockup_detector_event_create())
 190                 return;
 191
 192         if (!watchdog_cpus++)
 193                 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 194
 195         perf_event_enable(this_cpu_read(watchdog_ev));
 196 }
 197
 198 /**
 199  * hardlockup_detector_perf_disable - Disable the local event
 200  */
 201 void hardlockup_detector_perf_disable(void)
 202 {
 203         struct perf_event *event = this_cpu_read(watchdog_ev);
 204
 205         if (event) {
 206                 perf_event_disable(event);
 207                 cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
 208                 watchdog_cpus--;
 209         }
 210 }
 211
 212 /**
 213  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
 214  *
 215  * Called from lockup_detector_cleanup(). Serialized by the caller.
 216  */
 217 void hardlockup_detector_perf_cleanup(void)
 218 {
 219         int cpu;
 220
 221         for_each_cpu(cpu, &dead_events_mask) {
 222                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 223
 224                 /*
 225                  * Required because for_each_cpu() reports  unconditionally
 226                  * CPU0 as set on UP kernels. Sigh.
 227                  */
 228                 if (event)
 229                         perf_event_release_kernel(event);
 230                 per_cpu(watchdog_ev, cpu) = NULL;
 231         }
 232         cpumask_clear(&dead_events_mask);
 233 }
 234
 235 /**
 236  * hardlockup_detector_perf_stop - Globally stop watchdog events
 237  *
 238  * Special interface for x86 to handle the perf HT bug.
 239  */
 240 void __init hardlockup_detector_perf_stop(void)
 241 {
 242         int cpu;
 243
 244         lockdep_assert_cpus_held();
 245
 246         for_each_online_cpu(cpu) {
 247                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 248
 249                 if (event)
 250                         perf_event_disable(event);
 251         }
 252 }
 253
 254 /**
 255  * hardlockup_detector_perf_restart - Globally restart watchdog events
 256  *
 257  * Special interface for x86 to handle the perf HT bug.
 258  */
 259 void __init hardlockup_detector_perf_restart(void)
 260 {
 261         int cpu;
 262
 263         lockdep_assert_cpus_held();
 264
 265         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 266                 return;
 267
 268         for_each_online_cpu(cpu) {
 269                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 270
 271                 if (event)
 272                         perf_event_enable(event);
 273         }
 274 }
 275
 276 /**
 277  * hardlockup_detector_perf_init - Probe whether NMI event is available at all
 278  */
 279 int __init hardlockup_detector_perf_init(void)
 280 {
 281         int ret = hardlockup_detector_event_create();
 282
 283         if (ret) {
 284                 pr_info("Perf NMI watchdog permanently disabled\n");
 285         } else {
 286                 perf_event_release_kernel(this_cpu_read(watchdog_ev));
 287                 this_cpu_write(watchdog_ev, NULL);
 288         }
 289         return ret;
 290 }