| 1 | /* |
| 2 | * Detect hard and soft lockups on a system |
| 3 | * |
| 4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. |
| 5 | * |
| 6 | * Note: Most of this code is borrowed heavily from the original softlockup |
| 7 | * detector, so thanks to Ingo for the initial implementation. |
| 8 | * Some chunks also taken from the old x86-specific nmi watchdog code, thanks |
| 9 | * to those contributors as well. |
| 10 | */ |
| 11 | |
| 12 | #define pr_fmt(fmt) "NMI watchdog: " fmt |
| 13 | |
| 14 | #include <linux/mm.h> |
| 15 | #include <linux/cpu.h> |
| 16 | #include <linux/nmi.h> |
| 17 | #include <linux/init.h> |
| 18 | #include <linux/delay.h> |
| 19 | #include <linux/freezer.h> |
| 20 | #include <linux/kthread.h> |
| 21 | #include <linux/lockdep.h> |
| 22 | #include <linux/notifier.h> |
| 23 | #include <linux/module.h> |
| 24 | #include <linux/sysctl.h> |
| 25 | #include <linux/smpboot.h> |
| 26 | #include <linux/sched/rt.h> |
| 27 | |
| 28 | #include <asm/irq_regs.h> |
| 29 | #include <linux/kvm_para.h> |
| 30 | #include <linux/perf_event.h> |
| 31 | |
| 32 | int watchdog_user_enabled = 1; |
| 33 | int __read_mostly watchdog_thresh = 10; |
| 34 | #ifdef CONFIG_SMP |
| 35 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
| 36 | #else |
| 37 | #define sysctl_softlockup_all_cpu_backtrace 0 |
| 38 | #endif |
| 39 | |
| 40 | static int __read_mostly watchdog_running; |
| 41 | static u64 __read_mostly sample_period; |
| 42 | |
| 43 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); |
| 44 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
| 45 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); |
| 46 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); |
| 47 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
| 48 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
| 49 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); |
| 50 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 51 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
| 52 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); |
| 53 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
| 54 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); |
| 55 | #endif |
| 56 | static unsigned long soft_lockup_nmi_warn; |
| 57 | |
| 58 | /* boot commands */ |
| 59 | /* |
| 60 | * Should we panic when a soft-lockup or hard-lockup occurs: |
| 61 | */ |
| 62 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 63 | static int hardlockup_panic = |
| 64 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; |
| 65 | |
| 66 | static int __init hardlockup_panic_setup(char *str) |
| 67 | { |
| 68 | if (!strncmp(str, "panic", 5)) |
| 69 | hardlockup_panic = 1; |
| 70 | else if (!strncmp(str, "nopanic", 7)) |
| 71 | hardlockup_panic = 0; |
| 72 | else if (!strncmp(str, "0", 1)) |
| 73 | watchdog_user_enabled = 0; |
| 74 | return 1; |
| 75 | } |
| 76 | __setup("nmi_watchdog=", hardlockup_panic_setup); |
| 77 | #endif |
| 78 | |
| 79 | unsigned int __read_mostly softlockup_panic = |
| 80 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; |
| 81 | |
| 82 | static int __init softlockup_panic_setup(char *str) |
| 83 | { |
| 84 | softlockup_panic = simple_strtoul(str, NULL, 0); |
| 85 | |
| 86 | return 1; |
| 87 | } |
| 88 | __setup("softlockup_panic=", softlockup_panic_setup); |
| 89 | |
| 90 | static int __init nowatchdog_setup(char *str) |
| 91 | { |
| 92 | watchdog_user_enabled = 0; |
| 93 | return 1; |
| 94 | } |
| 95 | __setup("nowatchdog", nowatchdog_setup); |
| 96 | |
| 97 | /* deprecated */ |
| 98 | static int __init nosoftlockup_setup(char *str) |
| 99 | { |
| 100 | watchdog_user_enabled = 0; |
| 101 | return 1; |
| 102 | } |
| 103 | __setup("nosoftlockup", nosoftlockup_setup); |
| 104 | /* */ |
| 105 | #ifdef CONFIG_SMP |
| 106 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
| 107 | { |
| 108 | sysctl_softlockup_all_cpu_backtrace = |
| 109 | !!simple_strtol(str, NULL, 0); |
| 110 | return 1; |
| 111 | } |
| 112 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); |
| 113 | #endif |
| 114 | |
| 115 | /* |
| 116 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- |
| 117 | * lockups can have false positives under extreme conditions. So we generally |
| 118 | * want a higher threshold for soft lockups than for hard lockups. So we couple |
| 119 | * the thresholds with a factor: we make the soft threshold twice the amount of |
| 120 | * time the hard threshold is. |
| 121 | */ |
| 122 | static int get_softlockup_thresh(void) |
| 123 | { |
| 124 | return watchdog_thresh * 2; |
| 125 | } |
| 126 | |
| 127 | /* |
| 128 | * Returns seconds, approximately. We don't need nanosecond |
| 129 | * resolution, and we don't need to waste time with a big divide when |
| 130 | * 2^30ns == 1.074s. |
| 131 | */ |
| 132 | static unsigned long get_timestamp(void) |
| 133 | { |
| 134 | return local_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
| 135 | } |
| 136 | |
| 137 | static void set_sample_period(void) |
| 138 | { |
| 139 | /* |
| 140 | * convert watchdog_thresh from seconds to ns |
| 141 | * the divide by 5 is to give hrtimer several chances (two |
| 142 | * or three with the current relation between the soft |
| 143 | * and hard thresholds) to increment before the |
| 144 | * hardlockup detector generates a warning |
| 145 | */ |
| 146 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
| 147 | } |
| 148 | |
| 149 | /* Commands for resetting the watchdog */ |
| 150 | static void __touch_watchdog(void) |
| 151 | { |
| 152 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
| 153 | } |
| 154 | |
| 155 | void touch_softlockup_watchdog(void) |
| 156 | { |
| 157 | /* |
| 158 | * Preemption can be enabled. It doesn't matter which CPU's timestamp |
| 159 | * gets zeroed here, so use the raw_ operation. |
| 160 | */ |
| 161 | raw_cpu_write(watchdog_touch_ts, 0); |
| 162 | } |
| 163 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
| 164 | |
| 165 | void touch_all_softlockup_watchdogs(void) |
| 166 | { |
| 167 | int cpu; |
| 168 | |
| 169 | /* |
| 170 | * this is done lockless |
| 171 | * do we care if a 0 races with a timestamp? |
| 172 | * all it means is the softlock check starts one cycle later |
| 173 | */ |
| 174 | for_each_online_cpu(cpu) |
| 175 | per_cpu(watchdog_touch_ts, cpu) = 0; |
| 176 | } |
| 177 | |
| 178 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 179 | void touch_nmi_watchdog(void) |
| 180 | { |
| 181 | /* |
| 182 | * Using __raw here because some code paths have |
| 183 | * preemption enabled. If preemption is enabled |
| 184 | * then interrupts should be enabled too, in which |
| 185 | * case we shouldn't have to worry about the watchdog |
| 186 | * going off. |
| 187 | */ |
| 188 | __raw_get_cpu_var(watchdog_nmi_touch) = true; |
| 189 | touch_softlockup_watchdog(); |
| 190 | } |
| 191 | EXPORT_SYMBOL(touch_nmi_watchdog); |
| 192 | |
| 193 | #endif |
| 194 | |
| 195 | void touch_softlockup_watchdog_sync(void) |
| 196 | { |
| 197 | __raw_get_cpu_var(softlockup_touch_sync) = true; |
| 198 | __raw_get_cpu_var(watchdog_touch_ts) = 0; |
| 199 | } |
| 200 | |
| 201 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 202 | /* watchdog detector functions */ |
| 203 | static int is_hardlockup(void) |
| 204 | { |
| 205 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
| 206 | |
| 207 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
| 208 | return 1; |
| 209 | |
| 210 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
| 211 | return 0; |
| 212 | } |
| 213 | #endif |
| 214 | |
| 215 | static int is_softlockup(unsigned long touch_ts) |
| 216 | { |
| 217 | unsigned long now = get_timestamp(); |
| 218 | |
| 219 | /* Warn about unreasonable delays: */ |
| 220 | if (time_after(now, touch_ts + get_softlockup_thresh())) |
| 221 | return now - touch_ts; |
| 222 | |
| 223 | return 0; |
| 224 | } |
| 225 | |
| 226 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 227 | |
| 228 | static struct perf_event_attr wd_hw_attr = { |
| 229 | .type = PERF_TYPE_HARDWARE, |
| 230 | .config = PERF_COUNT_HW_CPU_CYCLES, |
| 231 | .size = sizeof(struct perf_event_attr), |
| 232 | .pinned = 1, |
| 233 | .disabled = 1, |
| 234 | }; |
| 235 | |
| 236 | /* Callback function for perf event subsystem */ |
| 237 | static void watchdog_overflow_callback(struct perf_event *event, |
| 238 | struct perf_sample_data *data, |
| 239 | struct pt_regs *regs) |
| 240 | { |
| 241 | /* Ensure the watchdog never gets throttled */ |
| 242 | event->hw.interrupts = 0; |
| 243 | |
| 244 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
| 245 | __this_cpu_write(watchdog_nmi_touch, false); |
| 246 | return; |
| 247 | } |
| 248 | |
| 249 | /* check for a hardlockup |
| 250 | * This is done by making sure our timer interrupt |
| 251 | * is incrementing. The timer interrupt should have |
| 252 | * fired multiple times before we overflow'd. If it hasn't |
| 253 | * then this is a good indication the cpu is stuck |
| 254 | */ |
| 255 | if (is_hardlockup()) { |
| 256 | int this_cpu = smp_processor_id(); |
| 257 | |
| 258 | /* only print hardlockups once */ |
| 259 | if (__this_cpu_read(hard_watchdog_warn) == true) |
| 260 | return; |
| 261 | |
| 262 | if (hardlockup_panic) |
| 263 | panic("Watchdog detected hard LOCKUP on cpu %d", |
| 264 | this_cpu); |
| 265 | else |
| 266 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", |
| 267 | this_cpu); |
| 268 | |
| 269 | __this_cpu_write(hard_watchdog_warn, true); |
| 270 | return; |
| 271 | } |
| 272 | |
| 273 | __this_cpu_write(hard_watchdog_warn, false); |
| 274 | return; |
| 275 | } |
| 276 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
| 277 | |
| 278 | static void watchdog_interrupt_count(void) |
| 279 | { |
| 280 | __this_cpu_inc(hrtimer_interrupts); |
| 281 | } |
| 282 | |
| 283 | static int watchdog_nmi_enable(unsigned int cpu); |
| 284 | static void watchdog_nmi_disable(unsigned int cpu); |
| 285 | |
| 286 | /* watchdog kicker functions */ |
| 287 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) |
| 288 | { |
| 289 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
| 290 | struct pt_regs *regs = get_irq_regs(); |
| 291 | int duration; |
| 292 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
| 293 | |
| 294 | /* kick the hardlockup detector */ |
| 295 | watchdog_interrupt_count(); |
| 296 | |
| 297 | /* kick the softlockup detector */ |
| 298 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
| 299 | |
| 300 | /* .. and repeat */ |
| 301 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
| 302 | |
| 303 | if (touch_ts == 0) { |
| 304 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
| 305 | /* |
| 306 | * If the time stamp was touched atomically |
| 307 | * make sure the scheduler tick is up to date. |
| 308 | */ |
| 309 | __this_cpu_write(softlockup_touch_sync, false); |
| 310 | sched_clock_tick(); |
| 311 | } |
| 312 | |
| 313 | /* Clear the guest paused flag on watchdog reset */ |
| 314 | kvm_check_and_clear_guest_paused(); |
| 315 | __touch_watchdog(); |
| 316 | return HRTIMER_RESTART; |
| 317 | } |
| 318 | |
| 319 | /* check for a softlockup |
| 320 | * This is done by making sure a high priority task is |
| 321 | * being scheduled. The task touches the watchdog to |
| 322 | * indicate it is getting cpu time. If it hasn't then |
| 323 | * this is a good indication some task is hogging the cpu |
| 324 | */ |
| 325 | duration = is_softlockup(touch_ts); |
| 326 | if (unlikely(duration)) { |
| 327 | /* |
| 328 | * If a virtual machine is stopped by the host it can look to |
| 329 | * the watchdog like a soft lockup, check to see if the host |
| 330 | * stopped the vm before we issue the warning |
| 331 | */ |
| 332 | if (kvm_check_and_clear_guest_paused()) |
| 333 | return HRTIMER_RESTART; |
| 334 | |
| 335 | /* only warn once */ |
| 336 | if (__this_cpu_read(soft_watchdog_warn) == true) |
| 337 | return HRTIMER_RESTART; |
| 338 | |
| 339 | if (softlockup_all_cpu_backtrace) { |
| 340 | /* Prevent multiple soft-lockup reports if one cpu is already |
| 341 | * engaged in dumping cpu back traces |
| 342 | */ |
| 343 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { |
| 344 | /* Someone else will report us. Let's give up */ |
| 345 | __this_cpu_write(soft_watchdog_warn, true); |
| 346 | return HRTIMER_RESTART; |
| 347 | } |
| 348 | } |
| 349 | |
| 350 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
| 351 | smp_processor_id(), duration, |
| 352 | current->comm, task_pid_nr(current)); |
| 353 | print_modules(); |
| 354 | print_irqtrace_events(current); |
| 355 | if (regs) |
| 356 | show_regs(regs); |
| 357 | else |
| 358 | dump_stack(); |
| 359 | |
| 360 | if (softlockup_all_cpu_backtrace) { |
| 361 | /* Avoid generating two back traces for current |
| 362 | * given that one is already made above |
| 363 | */ |
| 364 | trigger_allbutself_cpu_backtrace(); |
| 365 | |
| 366 | clear_bit(0, &soft_lockup_nmi_warn); |
| 367 | /* Barrier to sync with other cpus */ |
| 368 | smp_mb__after_atomic(); |
| 369 | } |
| 370 | |
| 371 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
| 372 | if (softlockup_panic) |
| 373 | panic("softlockup: hung tasks"); |
| 374 | __this_cpu_write(soft_watchdog_warn, true); |
| 375 | } else |
| 376 | __this_cpu_write(soft_watchdog_warn, false); |
| 377 | |
| 378 | return HRTIMER_RESTART; |
| 379 | } |
| 380 | |
| 381 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
| 382 | { |
| 383 | struct sched_param param = { .sched_priority = prio }; |
| 384 | |
| 385 | sched_setscheduler(current, policy, ¶m); |
| 386 | } |
| 387 | |
| 388 | static void watchdog_enable(unsigned int cpu) |
| 389 | { |
| 390 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); |
| 391 | |
| 392 | /* kick off the timer for the hardlockup detector */ |
| 393 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 394 | hrtimer->function = watchdog_timer_fn; |
| 395 | |
| 396 | /* Enable the perf event */ |
| 397 | watchdog_nmi_enable(cpu); |
| 398 | |
| 399 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
| 400 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
| 401 | HRTIMER_MODE_REL_PINNED); |
| 402 | |
| 403 | /* initialize timestamp */ |
| 404 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
| 405 | __touch_watchdog(); |
| 406 | } |
| 407 | |
| 408 | static void watchdog_disable(unsigned int cpu) |
| 409 | { |
| 410 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); |
| 411 | |
| 412 | watchdog_set_prio(SCHED_NORMAL, 0); |
| 413 | hrtimer_cancel(hrtimer); |
| 414 | /* disable the perf event */ |
| 415 | watchdog_nmi_disable(cpu); |
| 416 | } |
| 417 | |
| 418 | static void watchdog_cleanup(unsigned int cpu, bool online) |
| 419 | { |
| 420 | watchdog_disable(cpu); |
| 421 | } |
| 422 | |
| 423 | static int watchdog_should_run(unsigned int cpu) |
| 424 | { |
| 425 | return __this_cpu_read(hrtimer_interrupts) != |
| 426 | __this_cpu_read(soft_lockup_hrtimer_cnt); |
| 427 | } |
| 428 | |
| 429 | /* |
| 430 | * The watchdog thread function - touches the timestamp. |
| 431 | * |
| 432 | * It only runs once every sample_period seconds (4 seconds by |
| 433 | * default) to reset the softlockup timestamp. If this gets delayed |
| 434 | * for more than 2*watchdog_thresh seconds then the debug-printout |
| 435 | * triggers in watchdog_timer_fn(). |
| 436 | */ |
| 437 | static void watchdog(unsigned int cpu) |
| 438 | { |
| 439 | __this_cpu_write(soft_lockup_hrtimer_cnt, |
| 440 | __this_cpu_read(hrtimer_interrupts)); |
| 441 | __touch_watchdog(); |
| 442 | } |
| 443 | |
| 444 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
| 445 | /* |
| 446 | * People like the simple clean cpu node info on boot. |
| 447 | * Reduce the watchdog noise by only printing messages |
| 448 | * that are different from what cpu0 displayed. |
| 449 | */ |
| 450 | static unsigned long cpu0_err; |
| 451 | |
| 452 | static int watchdog_nmi_enable(unsigned int cpu) |
| 453 | { |
| 454 | struct perf_event_attr *wd_attr; |
| 455 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
| 456 | |
| 457 | /* is it already setup and enabled? */ |
| 458 | if (event && event->state > PERF_EVENT_STATE_OFF) |
| 459 | goto out; |
| 460 | |
| 461 | /* it is setup but not enabled */ |
| 462 | if (event != NULL) |
| 463 | goto out_enable; |
| 464 | |
| 465 | wd_attr = &wd_hw_attr; |
| 466 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); |
| 467 | |
| 468 | /* Try to register using hardware perf events */ |
| 469 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); |
| 470 | |
| 471 | /* save cpu0 error for future comparision */ |
| 472 | if (cpu == 0 && IS_ERR(event)) |
| 473 | cpu0_err = PTR_ERR(event); |
| 474 | |
| 475 | if (!IS_ERR(event)) { |
| 476 | /* only print for cpu0 or different than cpu0 */ |
| 477 | if (cpu == 0 || cpu0_err) |
| 478 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); |
| 479 | goto out_save; |
| 480 | } |
| 481 | |
| 482 | /* skip displaying the same error again */ |
| 483 | if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) |
| 484 | return PTR_ERR(event); |
| 485 | |
| 486 | /* vary the KERN level based on the returned errno */ |
| 487 | if (PTR_ERR(event) == -EOPNOTSUPP) |
| 488 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); |
| 489 | else if (PTR_ERR(event) == -ENOENT) |
| 490 | pr_warn("disabled (cpu%i): hardware events not enabled\n", |
| 491 | cpu); |
| 492 | else |
| 493 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", |
| 494 | cpu, PTR_ERR(event)); |
| 495 | return PTR_ERR(event); |
| 496 | |
| 497 | /* success path */ |
| 498 | out_save: |
| 499 | per_cpu(watchdog_ev, cpu) = event; |
| 500 | out_enable: |
| 501 | perf_event_enable(per_cpu(watchdog_ev, cpu)); |
| 502 | out: |
| 503 | return 0; |
| 504 | } |
| 505 | |
| 506 | static void watchdog_nmi_disable(unsigned int cpu) |
| 507 | { |
| 508 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
| 509 | |
| 510 | if (event) { |
| 511 | perf_event_disable(event); |
| 512 | per_cpu(watchdog_ev, cpu) = NULL; |
| 513 | |
| 514 | /* should be in cleanup, but blocks oprofile */ |
| 515 | perf_event_release_kernel(event); |
| 516 | } |
| 517 | return; |
| 518 | } |
| 519 | #else |
| 520 | static int watchdog_nmi_enable(unsigned int cpu) { return 0; } |
| 521 | static void watchdog_nmi_disable(unsigned int cpu) { return; } |
| 522 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
| 523 | |
| 524 | static struct smp_hotplug_thread watchdog_threads = { |
| 525 | .store = &softlockup_watchdog, |
| 526 | .thread_should_run = watchdog_should_run, |
| 527 | .thread_fn = watchdog, |
| 528 | .thread_comm = "watchdog/%u", |
| 529 | .setup = watchdog_enable, |
| 530 | .cleanup = watchdog_cleanup, |
| 531 | .park = watchdog_disable, |
| 532 | .unpark = watchdog_enable, |
| 533 | }; |
| 534 | |
| 535 | static void restart_watchdog_hrtimer(void *info) |
| 536 | { |
| 537 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); |
| 538 | int ret; |
| 539 | |
| 540 | /* |
| 541 | * No need to cancel and restart hrtimer if it is currently executing |
| 542 | * because it will reprogram itself with the new period now. |
| 543 | * We should never see it unqueued here because we are running per-cpu |
| 544 | * with interrupts disabled. |
| 545 | */ |
| 546 | ret = hrtimer_try_to_cancel(hrtimer); |
| 547 | if (ret == 1) |
| 548 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
| 549 | HRTIMER_MODE_REL_PINNED); |
| 550 | } |
| 551 | |
| 552 | static void update_timers(int cpu) |
| 553 | { |
| 554 | /* |
| 555 | * Make sure that perf event counter will adopt to a new |
| 556 | * sampling period. Updating the sampling period directly would |
| 557 | * be much nicer but we do not have an API for that now so |
| 558 | * let's use a big hammer. |
| 559 | * Hrtimer will adopt the new period on the next tick but this |
| 560 | * might be late already so we have to restart the timer as well. |
| 561 | */ |
| 562 | watchdog_nmi_disable(cpu); |
| 563 | smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1); |
| 564 | watchdog_nmi_enable(cpu); |
| 565 | } |
| 566 | |
| 567 | static void update_timers_all_cpus(void) |
| 568 | { |
| 569 | int cpu; |
| 570 | |
| 571 | get_online_cpus(); |
| 572 | for_each_online_cpu(cpu) |
| 573 | update_timers(cpu); |
| 574 | put_online_cpus(); |
| 575 | } |
| 576 | |
| 577 | static int watchdog_enable_all_cpus(bool sample_period_changed) |
| 578 | { |
| 579 | int err = 0; |
| 580 | |
| 581 | if (!watchdog_running) { |
| 582 | err = smpboot_register_percpu_thread(&watchdog_threads); |
| 583 | if (err) |
| 584 | pr_err("Failed to create watchdog threads, disabled\n"); |
| 585 | else |
| 586 | watchdog_running = 1; |
| 587 | } else if (sample_period_changed) { |
| 588 | update_timers_all_cpus(); |
| 589 | } |
| 590 | |
| 591 | return err; |
| 592 | } |
| 593 | |
| 594 | /* prepare/enable/disable routines */ |
| 595 | /* sysctl functions */ |
| 596 | #ifdef CONFIG_SYSCTL |
| 597 | static void watchdog_disable_all_cpus(void) |
| 598 | { |
| 599 | if (watchdog_running) { |
| 600 | watchdog_running = 0; |
| 601 | smpboot_unregister_percpu_thread(&watchdog_threads); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | /* |
| 606 | * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh |
| 607 | */ |
| 608 | |
| 609 | int proc_dowatchdog(struct ctl_table *table, int write, |
| 610 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 611 | { |
| 612 | int err, old_thresh, old_enabled; |
| 613 | static DEFINE_MUTEX(watchdog_proc_mutex); |
| 614 | |
| 615 | mutex_lock(&watchdog_proc_mutex); |
| 616 | old_thresh = ACCESS_ONCE(watchdog_thresh); |
| 617 | old_enabled = ACCESS_ONCE(watchdog_user_enabled); |
| 618 | |
| 619 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 620 | if (err || !write) |
| 621 | goto out; |
| 622 | |
| 623 | set_sample_period(); |
| 624 | /* |
| 625 | * Watchdog threads shouldn't be enabled if they are |
| 626 | * disabled. The 'watchdog_running' variable check in |
| 627 | * watchdog_*_all_cpus() function takes care of this. |
| 628 | */ |
| 629 | if (watchdog_user_enabled && watchdog_thresh) |
| 630 | err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh); |
| 631 | else |
| 632 | watchdog_disable_all_cpus(); |
| 633 | |
| 634 | /* Restore old values on failure */ |
| 635 | if (err) { |
| 636 | watchdog_thresh = old_thresh; |
| 637 | watchdog_user_enabled = old_enabled; |
| 638 | } |
| 639 | out: |
| 640 | mutex_unlock(&watchdog_proc_mutex); |
| 641 | return err; |
| 642 | } |
| 643 | #endif /* CONFIG_SYSCTL */ |
| 644 | |
| 645 | void __init lockup_detector_init(void) |
| 646 | { |
| 647 | set_sample_period(); |
| 648 | |
| 649 | if (watchdog_user_enabled) |
| 650 | watchdog_enable_all_cpus(false); |
| 651 | } |