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 #else 79 static inline bool watchdog_check_timestamp(void) 80 { 81 return true; 82 } 83 #endif 84 85 static struct perf_event_attr wd_hw_attr = { 86 .type = PERF_TYPE_HARDWARE, 87 .config = PERF_COUNT_HW_CPU_CYCLES, 88 .size = sizeof(struct perf_event_attr), 89 .pinned = 1, 90 .disabled = 1, 91 }; 92 93 static struct perf_event_attr fallback_wd_hw_attr = { 94 .type = PERF_TYPE_HARDWARE, 95 .config = PERF_COUNT_HW_CPU_CYCLES, 96 .size = sizeof(struct perf_event_attr), 97 .pinned = 1, 98 .disabled = 1, 99 }; 100 101 /* Callback function for perf event subsystem */ 102 static void watchdog_overflow_callback(struct perf_event *event, 103 struct perf_sample_data *data, 104 struct pt_regs *regs) 105 { 106 /* Ensure the watchdog never gets throttled */ 107 event->hw.interrupts = 0; 108 109 if (!watchdog_check_timestamp()) 110 return; 111 112 watchdog_hardlockup_check(smp_processor_id(), regs); 113 } 114 115 static int hardlockup_detector_event_create(void) 116 { 117 unsigned int cpu; 118 struct perf_event_attr *wd_attr; 119 struct perf_event *evt; 120 121 /* 122 * Preemption is not disabled because memory will be allocated. 123 * Ensure CPU-locality by calling this in per-CPU kthread. 124 */ 125 WARN_ON(!is_percpu_thread()); 126 cpu = raw_smp_processor_id(); 127 wd_attr = &wd_hw_attr; 128 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); 129 130 /* Try to register using hardware perf events */ 131 evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL, 132 watchdog_overflow_callback, NULL); 133 if (IS_ERR(evt)) { 134 wd_attr = &fallback_wd_hw_attr; 135 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); 136 evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL, 137 watchdog_overflow_callback, NULL); 138 } 139 140 if (IS_ERR(evt)) { 141 pr_debug("Perf event create on CPU %d failed with %ld\n", cpu, 142 PTR_ERR(evt)); 143 return PTR_ERR(evt); 144 } 145 this_cpu_write(watchdog_ev, evt); 146 return 0; 147 } 148 149 /** 150 * watchdog_hardlockup_enable - Enable the local event 151 * @cpu: The CPU to enable hard lockup on. 152 */ 153 void watchdog_hardlockup_enable(unsigned int cpu) 154 { 155 WARN_ON_ONCE(cpu != smp_processor_id()); 156 157 if (hardlockup_detector_event_create()) 158 return; 159 160 /* use original value for check */ 161 if (!atomic_fetch_inc(&watchdog_cpus)) 162 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n"); 163 164 perf_event_enable(this_cpu_read(watchdog_ev)); 165 } 166 167 /** 168 * watchdog_hardlockup_disable - Disable the local event 169 * @cpu: The CPU to enable hard lockup on. 170 */ 171 void watchdog_hardlockup_disable(unsigned int cpu) 172 { 173 struct perf_event *event = this_cpu_read(watchdog_ev); 174 175 WARN_ON_ONCE(cpu != smp_processor_id()); 176 177 if (event) { 178 perf_event_disable(event); 179 this_cpu_write(watchdog_ev, NULL); 180 this_cpu_write(dead_event, event); 181 cpumask_set_cpu(smp_processor_id(), &dead_events_mask); 182 atomic_dec(&watchdog_cpus); 183 } 184 } 185 186 /** 187 * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them 188 * 189 * Called from lockup_detector_cleanup(). Serialized by the caller. 190 */ 191 void hardlockup_detector_perf_cleanup(void) 192 { 193 int cpu; 194 195 for_each_cpu(cpu, &dead_events_mask) { 196 struct perf_event *event = per_cpu(dead_event, cpu); 197 198 /* 199 * Required because for_each_cpu() reports unconditionally 200 * CPU0 as set on UP kernels. Sigh. 201 */ 202 if (event) 203 perf_event_release_kernel(event); 204 per_cpu(dead_event, cpu) = NULL; 205 } 206 cpumask_clear(&dead_events_mask); 207 } 208 209 /** 210 * hardlockup_detector_perf_stop - Globally stop watchdog events 211 * 212 * Special interface for x86 to handle the perf HT bug. 213 */ 214 void __init hardlockup_detector_perf_stop(void) 215 { 216 int cpu; 217 218 lockdep_assert_cpus_held(); 219 220 for_each_online_cpu(cpu) { 221 struct perf_event *event = per_cpu(watchdog_ev, cpu); 222 223 if (event) 224 perf_event_disable(event); 225 } 226 } 227 228 /** 229 * hardlockup_detector_perf_restart - Globally restart watchdog events 230 * 231 * Special interface for x86 to handle the perf HT bug. 232 */ 233 void __init hardlockup_detector_perf_restart(void) 234 { 235 int cpu; 236 237 lockdep_assert_cpus_held(); 238 239 if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)) 240 return; 241 242 for_each_online_cpu(cpu) { 243 struct perf_event *event = per_cpu(watchdog_ev, cpu); 244 245 if (event) 246 perf_event_enable(event); 247 } 248 } 249 250 bool __weak __init arch_perf_nmi_is_available(void) 251 { 252 return true; 253 } 254 255 /** 256 * watchdog_hardlockup_probe - Probe whether NMI event is available at all 257 */ 258 int __init watchdog_hardlockup_probe(void) 259 { 260 int ret; 261 262 if (!arch_perf_nmi_is_available()) 263 return -ENODEV; 264 265 ret = hardlockup_detector_event_create(); 266 267 if (ret) { 268 pr_info("Perf NMI watchdog permanently disabled\n"); 269 } else { 270 perf_event_release_kernel(this_cpu_read(watchdog_ev)); 271 this_cpu_write(watchdog_ev, NULL); 272 } 273 return ret; 274 } 275 276 /** 277 * hardlockup_config_perf_event - Overwrite config of wd_hw_attr. 278 * @str: number which identifies the raw perf event to use 279 */ 280 void __init hardlockup_config_perf_event(const char *str) 281 { 282 u64 config; 283 char buf[24]; 284 char *comma = strchr(str, ','); 285 286 if (!comma) { 287 if (kstrtoull(str, 16, &config)) 288 return; 289 } else { 290 unsigned int len = comma - str; 291 292 if (len >= sizeof(buf)) 293 return; 294 295 if (strscpy(buf, str, sizeof(buf)) < 0) 296 return; 297 buf[len] = 0; 298 if (kstrtoull(buf, 16, &config)) 299 return; 300 } 301 302 wd_hw_attr.type = PERF_TYPE_RAW; 303 wd_hw_attr.config = config; 304 } 305