kernel/watchdog_perf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Detect hard lockups on a system using perf
   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(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 atomic_t watchdog_cpus = ATOMIC_INIT(0);
  28
  29 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
  30 static DEFINE_PER_CPU(ktime_t, last_timestamp);
  31 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
  32 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
  33
  34 void watchdog_update_hrtimer_threshold(u64 period)
  35 {
  36         /*
  37          * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
  38          *
  39          * So it runs effectively with 2.5 times the rate of the NMI
  40          * watchdog. That means the hrtimer should fire 2-3 times before
  41          * the NMI watchdog expires. The NMI watchdog on x86 is based on
  42          * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
  43          * might run way faster than expected and the NMI fires in a
  44          * smaller period than the one deduced from the nominal CPU
  45          * frequency. Depending on the Turbo-Mode factor this might be fast
  46          * enough to get the NMI period smaller than the hrtimer watchdog
  47          * period and trigger false positives.
  48          *
  49          * The sample threshold is used to check in the NMI handler whether
  50          * the minimum time between two NMI samples has elapsed. That
  51          * prevents false positives.
  52          *
  53          * Set this to 4/5 of the actual watchdog threshold period so the
  54          * hrtimer is guaranteed to fire at least once within the real
  55          * watchdog threshold.
  56          */
  57         watchdog_hrtimer_sample_threshold = period * 2;
  58 }
  59
  60 static bool watchdog_check_timestamp(void)
  61 {
  62         ktime_t delta, now = ktime_get_mono_fast_ns();
  63
  64         delta = now - __this_cpu_read(last_timestamp);
  65         if (delta < watchdog_hrtimer_sample_threshold) {
  66                 /*
  67                  * If ktime is jiffies based, a stalled timer would prevent
  68                  * jiffies from being incremented and the filter would look
  69                  * at a stale timestamp and never trigger.
  70                  */
  71                 if (__this_cpu_inc_return(nmi_rearmed) < 10)
  72                         return false;
  73         }
  74         __this_cpu_write(nmi_rearmed, 0);
  75         __this_cpu_write(last_timestamp, now);
  76         return true;
  77 }
  78
  79 static void watchdog_init_timestamp(void)
  80 {
  81         __this_cpu_write(nmi_rearmed, 0);
  82         __this_cpu_write(last_timestamp, ktime_get_mono_fast_ns());
  83 }
  84 #else
  85 static inline bool watchdog_check_timestamp(void) { return true; }
  86 static inline void watchdog_init_timestamp(void) { }
  87 #endif
  88
  89 static struct perf_event_attr wd_hw_attr = {
  90         .type           = PERF_TYPE_HARDWARE,
  91         .config         = PERF_COUNT_HW_CPU_CYCLES,
  92         .size           = sizeof(struct perf_event_attr),
  93         .pinned         = 1,
  94         .disabled       = 1,
  95 };
  96
  97 static struct perf_event_attr fallback_wd_hw_attr = {
  98         .type           = PERF_TYPE_HARDWARE,
  99         .config         = PERF_COUNT_HW_CPU_CYCLES,
 100         .size           = sizeof(struct perf_event_attr),
 101         .pinned         = 1,
 102         .disabled       = 1,
 103 };
 104
 105 /* Callback function for perf event subsystem */
 106 static void watchdog_overflow_callback(struct perf_event *event,
 107                                        struct perf_sample_data *data,
 108                                        struct pt_regs *regs)
 109 {
 110         /* Ensure the watchdog never gets throttled */
 111         event->hw.interrupts = 0;
 112
 113         if (!watchdog_check_timestamp())
 114                 return;
 115
 116         watchdog_hardlockup_check(smp_processor_id(), regs);
 117 }
 118
 119 static int hardlockup_detector_event_create(void)
 120 {
 121         unsigned int cpu;
 122         struct perf_event_attr *wd_attr;
 123         struct perf_event *evt;
 124
 125         /*
 126          * Preemption is not disabled because memory will be allocated.
 127          * Ensure CPU-locality by calling this in per-CPU kthread.
 128          */
 129         WARN_ON(!is_percpu_thread());
 130         cpu = raw_smp_processor_id();
 131         wd_attr = &wd_hw_attr;
 132         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 133
 134         /* Try to register using hardware perf events */
 135         evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
 136                                                watchdog_overflow_callback, NULL);
 137         if (IS_ERR(evt)) {
 138                 wd_attr = &fallback_wd_hw_attr;
 139                 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 140                 evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
 141                                                        watchdog_overflow_callback, NULL);
 142         }
 143
 144         if (IS_ERR(evt)) {
 145                 pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
 146                          PTR_ERR(evt));
 147                 return PTR_ERR(evt);
 148         }
 149         this_cpu_write(watchdog_ev, evt);
 150         return 0;
 151 }
 152
 153 /**
 154  * watchdog_hardlockup_enable - Enable the local event
 155  * @cpu: The CPU to enable hard lockup on.
 156  */
 157 void watchdog_hardlockup_enable(unsigned int cpu)
 158 {
 159         WARN_ON_ONCE(cpu != smp_processor_id());
 160
 161         if (hardlockup_detector_event_create())
 162                 return;
 163
 164         /* use original value for check */
 165         if (!atomic_fetch_inc(&watchdog_cpus))
 166                 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 167
 168         watchdog_init_timestamp();
 169         perf_event_enable(this_cpu_read(watchdog_ev));
 170 }
 171
 172 /**
 173  * watchdog_hardlockup_disable - Disable the local event
 174  * @cpu: The CPU to enable hard lockup on.
 175  */
 176 void watchdog_hardlockup_disable(unsigned int cpu)
 177 {
 178         struct perf_event *event = this_cpu_read(watchdog_ev);
 179
 180         WARN_ON_ONCE(cpu != smp_processor_id());
 181
 182         if (event) {
 183                 perf_event_disable(event);
 184                 this_cpu_write(watchdog_ev, NULL);
 185                 this_cpu_write(dead_event, event);
 186                 cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
 187                 atomic_dec(&watchdog_cpus);
 188         }
 189 }
 190
 191 /**
 192  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
 193  *
 194  * Called from lockup_detector_cleanup(). Serialized by the caller.
 195  */
 196 void hardlockup_detector_perf_cleanup(void)
 197 {
 198         int cpu;
 199
 200         for_each_cpu(cpu, &dead_events_mask) {
 201                 struct perf_event *event = per_cpu(dead_event, cpu);
 202
 203                 /*
 204                  * Required because for_each_cpu() reports  unconditionally
 205                  * CPU0 as set on UP kernels. Sigh.
 206                  */
 207                 if (event)
 208                         perf_event_release_kernel(event);
 209                 per_cpu(dead_event, cpu) = NULL;
 210         }
 211         cpumask_clear(&dead_events_mask);
 212 }
 213
 214 /**
 215  * hardlockup_detector_perf_stop - Globally stop watchdog events
 216  *
 217  * Special interface for x86 to handle the perf HT bug.
 218  */
 219 void __init hardlockup_detector_perf_stop(void)
 220 {
 221         int cpu;
 222
 223         lockdep_assert_cpus_held();
 224
 225         for_each_online_cpu(cpu) {
 226                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 227
 228                 if (event)
 229                         perf_event_disable(event);
 230         }
 231 }
 232
 233 /**
 234  * hardlockup_detector_perf_restart - Globally restart watchdog events
 235  *
 236  * Special interface for x86 to handle the perf HT bug.
 237  */
 238 void __init hardlockup_detector_perf_restart(void)
 239 {
 240         int cpu;
 241
 242         lockdep_assert_cpus_held();
 243
 244         if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
 245                 return;
 246
 247         for_each_online_cpu(cpu) {
 248                 struct perf_event *event = per_cpu(watchdog_ev, cpu);
 249
 250                 if (event)
 251                         perf_event_enable(event);
 252         }
 253 }
 254
 255 bool __weak __init arch_perf_nmi_is_available(void)
 256 {
 257         return true;
 258 }
 259
 260 /**
 261  * watchdog_hardlockup_probe - Probe whether NMI event is available at all
 262  */
 263 int __init watchdog_hardlockup_probe(void)
 264 {
 265         int ret;
 266
 267         if (!arch_perf_nmi_is_available())
 268                 return -ENODEV;
 269
 270         ret = hardlockup_detector_event_create();
 271
 272         if (ret) {
 273                 pr_info("Perf NMI watchdog permanently disabled\n");
 274         } else {
 275                 perf_event_release_kernel(this_cpu_read(watchdog_ev));
 276                 this_cpu_write(watchdog_ev, NULL);
 277         }
 278         return ret;
 279 }
 280
 281 /**
 282  * hardlockup_config_perf_event - Overwrite config of wd_hw_attr.
 283  * @str: number which identifies the raw perf event to use
 284  */
 285 void __init hardlockup_config_perf_event(const char *str)
 286 {
 287         u64 config;
 288         char buf[24];
 289         char *comma = strchr(str, ',');
 290
 291         if (!comma) {
 292                 if (kstrtoull(str, 16, &config))
 293                         return;
 294         } else {
 295                 unsigned int len = comma - str;
 296
 297                 if (len >= sizeof(buf))
 298                         return;
 299
 300                 if (strscpy(buf, str, sizeof(buf)) < 0)
 301                         return;
 302                 buf[len] = 0;
 303                 if (kstrtoull(buf, 16, &config))
 304                         return;
 305         }
 306
 307         wd_hw_attr.type = PERF_TYPE_RAW;
 308         wd_hw_attr.config = config;
 309 }