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 /* Callback function for perf event subsystem */ 98 static void watchdog_overflow_callback(struct perf_event *event, 99 struct perf_sample_data *data, 100 struct pt_regs *regs) 101 { 102 /* Ensure the watchdog never gets throttled */ 103 event->hw.interrupts = 0; 104 105 if (!watchdog_check_timestamp()) 106 return; 107 108 watchdog_hardlockup_check(smp_processor_id(), regs); 109 } 110 111 static int hardlockup_detector_event_create(void) 112 { 113 unsigned int cpu; 114 struct perf_event_attr *wd_attr; 115 struct perf_event *evt; 116 117 /* 118 * Preemption is not disabled because memory will be allocated. 119 * Ensure CPU-locality by calling this in per-CPU kthread. 120 */ 121 WARN_ON(!is_percpu_thread()); 122 cpu = raw_smp_processor_id(); 123 wd_attr = &wd_hw_attr; 124 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); 125 126 /* Try to register using hardware perf events */ 127 evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL, 128 watchdog_overflow_callback, NULL); 129 if (IS_ERR(evt)) { 130 pr_debug("Perf event create on CPU %d failed with %ld\n", cpu, 131 PTR_ERR(evt)); 132 return PTR_ERR(evt); 133 } 134 this_cpu_write(watchdog_ev, evt); 135 return 0; 136 } 137 138 /** 139 * watchdog_hardlockup_enable - Enable the local event 140 * 141 * @cpu: The CPU to enable hard lockup on. 142 */ 143 void watchdog_hardlockup_enable(unsigned int cpu) 144 { 145 WARN_ON_ONCE(cpu != smp_processor_id()); 146 147 if (hardlockup_detector_event_create()) 148 return; 149 150 /* use original value for check */ 151 if (!atomic_fetch_inc(&watchdog_cpus)) 152 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n"); 153 154 watchdog_init_timestamp(); 155 perf_event_enable(this_cpu_read(watchdog_ev)); 156 } 157 158 /** 159 * watchdog_hardlockup_disable - Disable the local event 160 * 161 * @cpu: The CPU to enable hard lockup on. 162 */ 163 void watchdog_hardlockup_disable(unsigned int cpu) 164 { 165 struct perf_event *event = this_cpu_read(watchdog_ev); 166 167 WARN_ON_ONCE(cpu != smp_processor_id()); 168 169 if (event) { 170 perf_event_disable(event); 171 this_cpu_write(watchdog_ev, NULL); 172 this_cpu_write(dead_event, event); 173 cpumask_set_cpu(smp_processor_id(), &dead_events_mask); 174 atomic_dec(&watchdog_cpus); 175 } 176 } 177 178 /** 179 * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them 180 * 181 * Called from lockup_detector_cleanup(). Serialized by the caller. 182 */ 183 void hardlockup_detector_perf_cleanup(void) 184 { 185 int cpu; 186 187 for_each_cpu(cpu, &dead_events_mask) { 188 struct perf_event *event = per_cpu(dead_event, cpu); 189 190 /* 191 * Required because for_each_cpu() reports unconditionally 192 * CPU0 as set on UP kernels. Sigh. 193 */ 194 if (event) 195 perf_event_release_kernel(event); 196 per_cpu(dead_event, cpu) = NULL; 197 } 198 cpumask_clear(&dead_events_mask); 199 } 200 201 /** 202 * hardlockup_detector_perf_stop - Globally stop watchdog events 203 * 204 * Special interface for x86 to handle the perf HT bug. 205 */ 206 void __init hardlockup_detector_perf_stop(void) 207 { 208 int cpu; 209 210 lockdep_assert_cpus_held(); 211 212 for_each_online_cpu(cpu) { 213 struct perf_event *event = per_cpu(watchdog_ev, cpu); 214 215 if (event) 216 perf_event_disable(event); 217 } 218 } 219 220 /** 221 * hardlockup_detector_perf_restart - Globally restart watchdog events 222 * 223 * Special interface for x86 to handle the perf HT bug. 224 */ 225 void __init hardlockup_detector_perf_restart(void) 226 { 227 int cpu; 228 229 lockdep_assert_cpus_held(); 230 231 if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)) 232 return; 233 234 for_each_online_cpu(cpu) { 235 struct perf_event *event = per_cpu(watchdog_ev, cpu); 236 237 if (event) 238 perf_event_enable(event); 239 } 240 } 241 242 bool __weak __init arch_perf_nmi_is_available(void) 243 { 244 return true; 245 } 246 247 /** 248 * watchdog_hardlockup_probe - Probe whether NMI event is available at all 249 */ 250 int __init watchdog_hardlockup_probe(void) 251 { 252 int ret; 253 254 if (!arch_perf_nmi_is_available()) 255 return -ENODEV; 256 257 ret = hardlockup_detector_event_create(); 258 259 if (ret) { 260 pr_info("Perf NMI watchdog permanently disabled\n"); 261 } else { 262 perf_event_release_kernel(this_cpu_read(watchdog_ev)); 263 this_cpu_write(watchdog_ev, NULL); 264 } 265 return ret; 266 } 267