| 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/module.h> |
| 19 | #include <linux/sysctl.h> |
| 20 | #include <linux/smpboot.h> |
| 21 | #include <linux/sched/rt.h> |
| 22 | #include <uapi/linux/sched/types.h> |
| 23 | #include <linux/tick.h> |
| 24 | #include <linux/workqueue.h> |
| 25 | #include <linux/sched/clock.h> |
| 26 | #include <linux/sched/debug.h> |
| 27 | |
| 28 | #include <asm/irq_regs.h> |
| 29 | #include <linux/kvm_para.h> |
| 30 | #include <linux/kthread.h> |
| 31 | |
| 32 | static DEFINE_MUTEX(watchdog_proc_mutex); |
| 33 | |
| 34 | #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR) |
| 35 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED; |
| 36 | #else |
| 37 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
| 38 | #endif |
| 39 | int __read_mostly nmi_watchdog_enabled; |
| 40 | int __read_mostly soft_watchdog_enabled; |
| 41 | int __read_mostly watchdog_user_enabled; |
| 42 | int __read_mostly watchdog_thresh = 10; |
| 43 | |
| 44 | #ifdef CONFIG_SMP |
| 45 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
| 46 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; |
| 47 | #endif |
| 48 | static struct cpumask watchdog_cpumask __read_mostly; |
| 49 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); |
| 50 | |
| 51 | /* Helper for online, unparked cpus. */ |
| 52 | #define for_each_watchdog_cpu(cpu) \ |
| 53 | for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask) |
| 54 | |
| 55 | atomic_t watchdog_park_in_progress = ATOMIC_INIT(0); |
| 56 | |
| 57 | /* |
| 58 | * The 'watchdog_running' variable is set to 1 when the watchdog threads |
| 59 | * are registered/started and is set to 0 when the watchdog threads are |
| 60 | * unregistered/stopped, so it is an indicator whether the threads exist. |
| 61 | */ |
| 62 | static int __read_mostly watchdog_running; |
| 63 | /* |
| 64 | * If a subsystem has a need to deactivate the watchdog temporarily, it |
| 65 | * can use the suspend/resume interface to achieve this. The content of |
| 66 | * the 'watchdog_suspended' variable reflects this state. Existing threads |
| 67 | * are parked/unparked by the lockup_detector_{suspend|resume} functions |
| 68 | * (see comment blocks pertaining to those functions for further details). |
| 69 | * |
| 70 | * 'watchdog_suspended' also prevents threads from being registered/started |
| 71 | * or unregistered/stopped via parameters in /proc/sys/kernel, so the state |
| 72 | * of 'watchdog_running' cannot change while the watchdog is deactivated |
| 73 | * temporarily (see related code in 'proc' handlers). |
| 74 | */ |
| 75 | static int __read_mostly watchdog_suspended; |
| 76 | |
| 77 | static u64 __read_mostly sample_period; |
| 78 | |
| 79 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); |
| 80 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
| 81 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); |
| 82 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); |
| 83 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
| 84 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
| 85 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); |
| 86 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
| 87 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
| 88 | static unsigned long soft_lockup_nmi_warn; |
| 89 | |
| 90 | unsigned int __read_mostly softlockup_panic = |
| 91 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; |
| 92 | |
| 93 | static int __init softlockup_panic_setup(char *str) |
| 94 | { |
| 95 | softlockup_panic = simple_strtoul(str, NULL, 0); |
| 96 | |
| 97 | return 1; |
| 98 | } |
| 99 | __setup("softlockup_panic=", softlockup_panic_setup); |
| 100 | |
| 101 | static int __init nowatchdog_setup(char *str) |
| 102 | { |
| 103 | watchdog_enabled = 0; |
| 104 | return 1; |
| 105 | } |
| 106 | __setup("nowatchdog", nowatchdog_setup); |
| 107 | |
| 108 | static int __init nosoftlockup_setup(char *str) |
| 109 | { |
| 110 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
| 111 | return 1; |
| 112 | } |
| 113 | __setup("nosoftlockup", nosoftlockup_setup); |
| 114 | |
| 115 | #ifdef CONFIG_SMP |
| 116 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
| 117 | { |
| 118 | sysctl_softlockup_all_cpu_backtrace = |
| 119 | !!simple_strtol(str, NULL, 0); |
| 120 | return 1; |
| 121 | } |
| 122 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); |
| 123 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
| 124 | { |
| 125 | sysctl_hardlockup_all_cpu_backtrace = |
| 126 | !!simple_strtol(str, NULL, 0); |
| 127 | return 1; |
| 128 | } |
| 129 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); |
| 130 | #endif |
| 131 | |
| 132 | /* |
| 133 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- |
| 134 | * lockups can have false positives under extreme conditions. So we generally |
| 135 | * want a higher threshold for soft lockups than for hard lockups. So we couple |
| 136 | * the thresholds with a factor: we make the soft threshold twice the amount of |
| 137 | * time the hard threshold is. |
| 138 | */ |
| 139 | static int get_softlockup_thresh(void) |
| 140 | { |
| 141 | return watchdog_thresh * 2; |
| 142 | } |
| 143 | |
| 144 | /* |
| 145 | * Returns seconds, approximately. We don't need nanosecond |
| 146 | * resolution, and we don't need to waste time with a big divide when |
| 147 | * 2^30ns == 1.074s. |
| 148 | */ |
| 149 | static unsigned long get_timestamp(void) |
| 150 | { |
| 151 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
| 152 | } |
| 153 | |
| 154 | static void set_sample_period(void) |
| 155 | { |
| 156 | /* |
| 157 | * convert watchdog_thresh from seconds to ns |
| 158 | * the divide by 5 is to give hrtimer several chances (two |
| 159 | * or three with the current relation between the soft |
| 160 | * and hard thresholds) to increment before the |
| 161 | * hardlockup detector generates a warning |
| 162 | */ |
| 163 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
| 164 | } |
| 165 | |
| 166 | /* Commands for resetting the watchdog */ |
| 167 | static void __touch_watchdog(void) |
| 168 | { |
| 169 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
| 170 | } |
| 171 | |
| 172 | /** |
| 173 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls |
| 174 | * |
| 175 | * Call when the scheduler may have stalled for legitimate reasons |
| 176 | * preventing the watchdog task from executing - e.g. the scheduler |
| 177 | * entering idle state. This should only be used for scheduler events. |
| 178 | * Use touch_softlockup_watchdog() for everything else. |
| 179 | */ |
| 180 | void touch_softlockup_watchdog_sched(void) |
| 181 | { |
| 182 | /* |
| 183 | * Preemption can be enabled. It doesn't matter which CPU's timestamp |
| 184 | * gets zeroed here, so use the raw_ operation. |
| 185 | */ |
| 186 | raw_cpu_write(watchdog_touch_ts, 0); |
| 187 | } |
| 188 | |
| 189 | void touch_softlockup_watchdog(void) |
| 190 | { |
| 191 | touch_softlockup_watchdog_sched(); |
| 192 | wq_watchdog_touch(raw_smp_processor_id()); |
| 193 | } |
| 194 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
| 195 | |
| 196 | void touch_all_softlockup_watchdogs(void) |
| 197 | { |
| 198 | int cpu; |
| 199 | |
| 200 | /* |
| 201 | * this is done lockless |
| 202 | * do we care if a 0 races with a timestamp? |
| 203 | * all it means is the softlock check starts one cycle later |
| 204 | */ |
| 205 | for_each_watchdog_cpu(cpu) |
| 206 | per_cpu(watchdog_touch_ts, cpu) = 0; |
| 207 | wq_watchdog_touch(-1); |
| 208 | } |
| 209 | |
| 210 | void touch_softlockup_watchdog_sync(void) |
| 211 | { |
| 212 | __this_cpu_write(softlockup_touch_sync, true); |
| 213 | __this_cpu_write(watchdog_touch_ts, 0); |
| 214 | } |
| 215 | |
| 216 | /* watchdog detector functions */ |
| 217 | bool is_hardlockup(void) |
| 218 | { |
| 219 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
| 220 | |
| 221 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
| 222 | return true; |
| 223 | |
| 224 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
| 225 | return false; |
| 226 | } |
| 227 | |
| 228 | static int is_softlockup(unsigned long touch_ts) |
| 229 | { |
| 230 | unsigned long now = get_timestamp(); |
| 231 | |
| 232 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
| 233 | /* Warn about unreasonable delays. */ |
| 234 | if (time_after(now, touch_ts + get_softlockup_thresh())) |
| 235 | return now - touch_ts; |
| 236 | } |
| 237 | return 0; |
| 238 | } |
| 239 | |
| 240 | static void watchdog_interrupt_count(void) |
| 241 | { |
| 242 | __this_cpu_inc(hrtimer_interrupts); |
| 243 | } |
| 244 | |
| 245 | /* |
| 246 | * These two functions are mostly architecture specific |
| 247 | * defining them as weak here. |
| 248 | */ |
| 249 | int __weak watchdog_nmi_enable(unsigned int cpu) |
| 250 | { |
| 251 | return 0; |
| 252 | } |
| 253 | void __weak watchdog_nmi_disable(unsigned int cpu) |
| 254 | { |
| 255 | } |
| 256 | |
| 257 | static int watchdog_enable_all_cpus(void); |
| 258 | static void watchdog_disable_all_cpus(void); |
| 259 | |
| 260 | /* watchdog kicker functions */ |
| 261 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) |
| 262 | { |
| 263 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
| 264 | struct pt_regs *regs = get_irq_regs(); |
| 265 | int duration; |
| 266 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
| 267 | |
| 268 | if (atomic_read(&watchdog_park_in_progress) != 0) |
| 269 | return HRTIMER_NORESTART; |
| 270 | |
| 271 | /* kick the hardlockup detector */ |
| 272 | watchdog_interrupt_count(); |
| 273 | |
| 274 | /* kick the softlockup detector */ |
| 275 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
| 276 | |
| 277 | /* .. and repeat */ |
| 278 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
| 279 | |
| 280 | if (touch_ts == 0) { |
| 281 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
| 282 | /* |
| 283 | * If the time stamp was touched atomically |
| 284 | * make sure the scheduler tick is up to date. |
| 285 | */ |
| 286 | __this_cpu_write(softlockup_touch_sync, false); |
| 287 | sched_clock_tick(); |
| 288 | } |
| 289 | |
| 290 | /* Clear the guest paused flag on watchdog reset */ |
| 291 | kvm_check_and_clear_guest_paused(); |
| 292 | __touch_watchdog(); |
| 293 | return HRTIMER_RESTART; |
| 294 | } |
| 295 | |
| 296 | /* check for a softlockup |
| 297 | * This is done by making sure a high priority task is |
| 298 | * being scheduled. The task touches the watchdog to |
| 299 | * indicate it is getting cpu time. If it hasn't then |
| 300 | * this is a good indication some task is hogging the cpu |
| 301 | */ |
| 302 | duration = is_softlockup(touch_ts); |
| 303 | if (unlikely(duration)) { |
| 304 | /* |
| 305 | * If a virtual machine is stopped by the host it can look to |
| 306 | * the watchdog like a soft lockup, check to see if the host |
| 307 | * stopped the vm before we issue the warning |
| 308 | */ |
| 309 | if (kvm_check_and_clear_guest_paused()) |
| 310 | return HRTIMER_RESTART; |
| 311 | |
| 312 | /* only warn once */ |
| 313 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
| 314 | /* |
| 315 | * When multiple processes are causing softlockups the |
| 316 | * softlockup detector only warns on the first one |
| 317 | * because the code relies on a full quiet cycle to |
| 318 | * re-arm. The second process prevents the quiet cycle |
| 319 | * and never gets reported. Use task pointers to detect |
| 320 | * this. |
| 321 | */ |
| 322 | if (__this_cpu_read(softlockup_task_ptr_saved) != |
| 323 | current) { |
| 324 | __this_cpu_write(soft_watchdog_warn, false); |
| 325 | __touch_watchdog(); |
| 326 | } |
| 327 | return HRTIMER_RESTART; |
| 328 | } |
| 329 | |
| 330 | if (softlockup_all_cpu_backtrace) { |
| 331 | /* Prevent multiple soft-lockup reports if one cpu is already |
| 332 | * engaged in dumping cpu back traces |
| 333 | */ |
| 334 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { |
| 335 | /* Someone else will report us. Let's give up */ |
| 336 | __this_cpu_write(soft_watchdog_warn, true); |
| 337 | return HRTIMER_RESTART; |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
| 342 | smp_processor_id(), duration, |
| 343 | current->comm, task_pid_nr(current)); |
| 344 | __this_cpu_write(softlockup_task_ptr_saved, current); |
| 345 | print_modules(); |
| 346 | print_irqtrace_events(current); |
| 347 | if (regs) |
| 348 | show_regs(regs); |
| 349 | else |
| 350 | dump_stack(); |
| 351 | |
| 352 | if (softlockup_all_cpu_backtrace) { |
| 353 | /* Avoid generating two back traces for current |
| 354 | * given that one is already made above |
| 355 | */ |
| 356 | trigger_allbutself_cpu_backtrace(); |
| 357 | |
| 358 | clear_bit(0, &soft_lockup_nmi_warn); |
| 359 | /* Barrier to sync with other cpus */ |
| 360 | smp_mb__after_atomic(); |
| 361 | } |
| 362 | |
| 363 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
| 364 | if (softlockup_panic) |
| 365 | panic("softlockup: hung tasks"); |
| 366 | __this_cpu_write(soft_watchdog_warn, true); |
| 367 | } else |
| 368 | __this_cpu_write(soft_watchdog_warn, false); |
| 369 | |
| 370 | return HRTIMER_RESTART; |
| 371 | } |
| 372 | |
| 373 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
| 374 | { |
| 375 | struct sched_param param = { .sched_priority = prio }; |
| 376 | |
| 377 | sched_setscheduler(current, policy, ¶m); |
| 378 | } |
| 379 | |
| 380 | static void watchdog_enable(unsigned int cpu) |
| 381 | { |
| 382 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
| 383 | |
| 384 | /* kick off the timer for the hardlockup detector */ |
| 385 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 386 | hrtimer->function = watchdog_timer_fn; |
| 387 | |
| 388 | /* Enable the perf event */ |
| 389 | watchdog_nmi_enable(cpu); |
| 390 | |
| 391 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
| 392 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
| 393 | HRTIMER_MODE_REL_PINNED); |
| 394 | |
| 395 | /* initialize timestamp */ |
| 396 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
| 397 | __touch_watchdog(); |
| 398 | } |
| 399 | |
| 400 | static void watchdog_disable(unsigned int cpu) |
| 401 | { |
| 402 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
| 403 | |
| 404 | watchdog_set_prio(SCHED_NORMAL, 0); |
| 405 | hrtimer_cancel(hrtimer); |
| 406 | /* disable the perf event */ |
| 407 | watchdog_nmi_disable(cpu); |
| 408 | } |
| 409 | |
| 410 | static void watchdog_cleanup(unsigned int cpu, bool online) |
| 411 | { |
| 412 | watchdog_disable(cpu); |
| 413 | } |
| 414 | |
| 415 | static int watchdog_should_run(unsigned int cpu) |
| 416 | { |
| 417 | return __this_cpu_read(hrtimer_interrupts) != |
| 418 | __this_cpu_read(soft_lockup_hrtimer_cnt); |
| 419 | } |
| 420 | |
| 421 | /* |
| 422 | * The watchdog thread function - touches the timestamp. |
| 423 | * |
| 424 | * It only runs once every sample_period seconds (4 seconds by |
| 425 | * default) to reset the softlockup timestamp. If this gets delayed |
| 426 | * for more than 2*watchdog_thresh seconds then the debug-printout |
| 427 | * triggers in watchdog_timer_fn(). |
| 428 | */ |
| 429 | static void watchdog(unsigned int cpu) |
| 430 | { |
| 431 | __this_cpu_write(soft_lockup_hrtimer_cnt, |
| 432 | __this_cpu_read(hrtimer_interrupts)); |
| 433 | __touch_watchdog(); |
| 434 | |
| 435 | /* |
| 436 | * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the |
| 437 | * failure path. Check for failures that can occur asynchronously - |
| 438 | * for example, when CPUs are on-lined - and shut down the hardware |
| 439 | * perf event on each CPU accordingly. |
| 440 | * |
| 441 | * The only non-obvious place this bit can be cleared is through |
| 442 | * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a |
| 443 | * pr_info here would be too noisy as it would result in a message |
| 444 | * every few seconds if the hardlockup was disabled but the softlockup |
| 445 | * enabled. |
| 446 | */ |
| 447 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) |
| 448 | watchdog_nmi_disable(cpu); |
| 449 | } |
| 450 | |
| 451 | static struct smp_hotplug_thread watchdog_threads = { |
| 452 | .store = &softlockup_watchdog, |
| 453 | .thread_should_run = watchdog_should_run, |
| 454 | .thread_fn = watchdog, |
| 455 | .thread_comm = "watchdog/%u", |
| 456 | .setup = watchdog_enable, |
| 457 | .cleanup = watchdog_cleanup, |
| 458 | .park = watchdog_disable, |
| 459 | .unpark = watchdog_enable, |
| 460 | }; |
| 461 | |
| 462 | /* |
| 463 | * park all watchdog threads that are specified in 'watchdog_cpumask' |
| 464 | * |
| 465 | * This function returns an error if kthread_park() of a watchdog thread |
| 466 | * fails. In this situation, the watchdog threads of some CPUs can already |
| 467 | * be parked and the watchdog threads of other CPUs can still be runnable. |
| 468 | * Callers are expected to handle this special condition as appropriate in |
| 469 | * their context. |
| 470 | * |
| 471 | * This function may only be called in a context that is protected against |
| 472 | * races with CPU hotplug - for example, via get_online_cpus(). |
| 473 | */ |
| 474 | static int watchdog_park_threads(void) |
| 475 | { |
| 476 | int cpu, ret = 0; |
| 477 | |
| 478 | atomic_set(&watchdog_park_in_progress, 1); |
| 479 | |
| 480 | for_each_watchdog_cpu(cpu) { |
| 481 | ret = kthread_park(per_cpu(softlockup_watchdog, cpu)); |
| 482 | if (ret) |
| 483 | break; |
| 484 | } |
| 485 | |
| 486 | atomic_set(&watchdog_park_in_progress, 0); |
| 487 | |
| 488 | return ret; |
| 489 | } |
| 490 | |
| 491 | /* |
| 492 | * unpark all watchdog threads that are specified in 'watchdog_cpumask' |
| 493 | * |
| 494 | * This function may only be called in a context that is protected against |
| 495 | * races with CPU hotplug - for example, via get_online_cpus(). |
| 496 | */ |
| 497 | static void watchdog_unpark_threads(void) |
| 498 | { |
| 499 | int cpu; |
| 500 | |
| 501 | for_each_watchdog_cpu(cpu) |
| 502 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); |
| 503 | } |
| 504 | |
| 505 | /* |
| 506 | * Suspend the hard and soft lockup detector by parking the watchdog threads. |
| 507 | */ |
| 508 | int lockup_detector_suspend(void) |
| 509 | { |
| 510 | int ret = 0; |
| 511 | |
| 512 | get_online_cpus(); |
| 513 | mutex_lock(&watchdog_proc_mutex); |
| 514 | /* |
| 515 | * Multiple suspend requests can be active in parallel (counted by |
| 516 | * the 'watchdog_suspended' variable). If the watchdog threads are |
| 517 | * running, the first caller takes care that they will be parked. |
| 518 | * The state of 'watchdog_running' cannot change while a suspend |
| 519 | * request is active (see related code in 'proc' handlers). |
| 520 | */ |
| 521 | if (watchdog_running && !watchdog_suspended) |
| 522 | ret = watchdog_park_threads(); |
| 523 | |
| 524 | if (ret == 0) |
| 525 | watchdog_suspended++; |
| 526 | else { |
| 527 | watchdog_disable_all_cpus(); |
| 528 | pr_err("Failed to suspend lockup detectors, disabled\n"); |
| 529 | watchdog_enabled = 0; |
| 530 | } |
| 531 | |
| 532 | mutex_unlock(&watchdog_proc_mutex); |
| 533 | |
| 534 | return ret; |
| 535 | } |
| 536 | |
| 537 | /* |
| 538 | * Resume the hard and soft lockup detector by unparking the watchdog threads. |
| 539 | */ |
| 540 | void lockup_detector_resume(void) |
| 541 | { |
| 542 | mutex_lock(&watchdog_proc_mutex); |
| 543 | |
| 544 | watchdog_suspended--; |
| 545 | /* |
| 546 | * The watchdog threads are unparked if they were previously running |
| 547 | * and if there is no more active suspend request. |
| 548 | */ |
| 549 | if (watchdog_running && !watchdog_suspended) |
| 550 | watchdog_unpark_threads(); |
| 551 | |
| 552 | mutex_unlock(&watchdog_proc_mutex); |
| 553 | put_online_cpus(); |
| 554 | } |
| 555 | |
| 556 | static int update_watchdog_all_cpus(void) |
| 557 | { |
| 558 | int ret; |
| 559 | |
| 560 | ret = watchdog_park_threads(); |
| 561 | if (ret) |
| 562 | return ret; |
| 563 | |
| 564 | watchdog_unpark_threads(); |
| 565 | |
| 566 | return 0; |
| 567 | } |
| 568 | |
| 569 | static int watchdog_enable_all_cpus(void) |
| 570 | { |
| 571 | int err = 0; |
| 572 | |
| 573 | if (!watchdog_running) { |
| 574 | err = smpboot_register_percpu_thread_cpumask(&watchdog_threads, |
| 575 | &watchdog_cpumask); |
| 576 | if (err) |
| 577 | pr_err("Failed to create watchdog threads, disabled\n"); |
| 578 | else |
| 579 | watchdog_running = 1; |
| 580 | } else { |
| 581 | /* |
| 582 | * Enable/disable the lockup detectors or |
| 583 | * change the sample period 'on the fly'. |
| 584 | */ |
| 585 | err = update_watchdog_all_cpus(); |
| 586 | |
| 587 | if (err) { |
| 588 | watchdog_disable_all_cpus(); |
| 589 | pr_err("Failed to update lockup detectors, disabled\n"); |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | if (err) |
| 594 | watchdog_enabled = 0; |
| 595 | |
| 596 | return err; |
| 597 | } |
| 598 | |
| 599 | static void watchdog_disable_all_cpus(void) |
| 600 | { |
| 601 | if (watchdog_running) { |
| 602 | watchdog_running = 0; |
| 603 | smpboot_unregister_percpu_thread(&watchdog_threads); |
| 604 | } |
| 605 | } |
| 606 | |
| 607 | #ifdef CONFIG_SYSCTL |
| 608 | |
| 609 | /* |
| 610 | * Update the run state of the lockup detectors. |
| 611 | */ |
| 612 | static int proc_watchdog_update(void) |
| 613 | { |
| 614 | int err = 0; |
| 615 | |
| 616 | /* |
| 617 | * Watchdog threads won't be started if they are already active. |
| 618 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes |
| 619 | * care of this. If those threads are already active, the sample |
| 620 | * period will be updated and the lockup detectors will be enabled |
| 621 | * or disabled 'on the fly'. |
| 622 | */ |
| 623 | if (watchdog_enabled && watchdog_thresh) |
| 624 | err = watchdog_enable_all_cpus(); |
| 625 | else |
| 626 | watchdog_disable_all_cpus(); |
| 627 | |
| 628 | return err; |
| 629 | |
| 630 | } |
| 631 | |
| 632 | /* |
| 633 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter |
| 634 | * |
| 635 | * caller | table->data points to | 'which' contains the flag(s) |
| 636 | * -------------------|-----------------------|----------------------------- |
| 637 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed |
| 638 | * | | with SOFT_WATCHDOG_ENABLED |
| 639 | * -------------------|-----------------------|----------------------------- |
| 640 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED |
| 641 | * -------------------|-----------------------|----------------------------- |
| 642 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED |
| 643 | */ |
| 644 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, |
| 645 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 646 | { |
| 647 | int err, old, new; |
| 648 | int *watchdog_param = (int *)table->data; |
| 649 | |
| 650 | get_online_cpus(); |
| 651 | mutex_lock(&watchdog_proc_mutex); |
| 652 | |
| 653 | if (watchdog_suspended) { |
| 654 | /* no parameter changes allowed while watchdog is suspended */ |
| 655 | err = -EAGAIN; |
| 656 | goto out; |
| 657 | } |
| 658 | |
| 659 | /* |
| 660 | * If the parameter is being read return the state of the corresponding |
| 661 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the |
| 662 | * run state of the lockup detectors. |
| 663 | */ |
| 664 | if (!write) { |
| 665 | *watchdog_param = (watchdog_enabled & which) != 0; |
| 666 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 667 | } else { |
| 668 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 669 | if (err) |
| 670 | goto out; |
| 671 | |
| 672 | /* |
| 673 | * There is a race window between fetching the current value |
| 674 | * from 'watchdog_enabled' and storing the new value. During |
| 675 | * this race window, watchdog_nmi_enable() can sneak in and |
| 676 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. |
| 677 | * The 'cmpxchg' detects this race and the loop retries. |
| 678 | */ |
| 679 | do { |
| 680 | old = watchdog_enabled; |
| 681 | /* |
| 682 | * If the parameter value is not zero set the |
| 683 | * corresponding bit(s), else clear it(them). |
| 684 | */ |
| 685 | if (*watchdog_param) |
| 686 | new = old | which; |
| 687 | else |
| 688 | new = old & ~which; |
| 689 | } while (cmpxchg(&watchdog_enabled, old, new) != old); |
| 690 | |
| 691 | /* |
| 692 | * Update the run state of the lockup detectors. There is _no_ |
| 693 | * need to check the value returned by proc_watchdog_update() |
| 694 | * and to restore the previous value of 'watchdog_enabled' as |
| 695 | * both lockup detectors are disabled if proc_watchdog_update() |
| 696 | * returns an error. |
| 697 | */ |
| 698 | if (old == new) |
| 699 | goto out; |
| 700 | |
| 701 | err = proc_watchdog_update(); |
| 702 | } |
| 703 | out: |
| 704 | mutex_unlock(&watchdog_proc_mutex); |
| 705 | put_online_cpus(); |
| 706 | return err; |
| 707 | } |
| 708 | |
| 709 | /* |
| 710 | * /proc/sys/kernel/watchdog |
| 711 | */ |
| 712 | int proc_watchdog(struct ctl_table *table, int write, |
| 713 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 714 | { |
| 715 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, |
| 716 | table, write, buffer, lenp, ppos); |
| 717 | } |
| 718 | |
| 719 | /* |
| 720 | * /proc/sys/kernel/nmi_watchdog |
| 721 | */ |
| 722 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
| 723 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 724 | { |
| 725 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, |
| 726 | table, write, buffer, lenp, ppos); |
| 727 | } |
| 728 | |
| 729 | /* |
| 730 | * /proc/sys/kernel/soft_watchdog |
| 731 | */ |
| 732 | int proc_soft_watchdog(struct ctl_table *table, int write, |
| 733 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 734 | { |
| 735 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, |
| 736 | table, write, buffer, lenp, ppos); |
| 737 | } |
| 738 | |
| 739 | /* |
| 740 | * /proc/sys/kernel/watchdog_thresh |
| 741 | */ |
| 742 | int proc_watchdog_thresh(struct ctl_table *table, int write, |
| 743 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 744 | { |
| 745 | int err, old, new; |
| 746 | |
| 747 | get_online_cpus(); |
| 748 | mutex_lock(&watchdog_proc_mutex); |
| 749 | |
| 750 | if (watchdog_suspended) { |
| 751 | /* no parameter changes allowed while watchdog is suspended */ |
| 752 | err = -EAGAIN; |
| 753 | goto out; |
| 754 | } |
| 755 | |
| 756 | old = ACCESS_ONCE(watchdog_thresh); |
| 757 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 758 | |
| 759 | if (err || !write) |
| 760 | goto out; |
| 761 | |
| 762 | /* |
| 763 | * Update the sample period. Restore on failure. |
| 764 | */ |
| 765 | new = ACCESS_ONCE(watchdog_thresh); |
| 766 | if (old == new) |
| 767 | goto out; |
| 768 | |
| 769 | set_sample_period(); |
| 770 | err = proc_watchdog_update(); |
| 771 | if (err) { |
| 772 | watchdog_thresh = old; |
| 773 | set_sample_period(); |
| 774 | } |
| 775 | out: |
| 776 | mutex_unlock(&watchdog_proc_mutex); |
| 777 | put_online_cpus(); |
| 778 | return err; |
| 779 | } |
| 780 | |
| 781 | /* |
| 782 | * The cpumask is the mask of possible cpus that the watchdog can run |
| 783 | * on, not the mask of cpus it is actually running on. This allows the |
| 784 | * user to specify a mask that will include cpus that have not yet |
| 785 | * been brought online, if desired. |
| 786 | */ |
| 787 | int proc_watchdog_cpumask(struct ctl_table *table, int write, |
| 788 | void __user *buffer, size_t *lenp, loff_t *ppos) |
| 789 | { |
| 790 | int err; |
| 791 | |
| 792 | get_online_cpus(); |
| 793 | mutex_lock(&watchdog_proc_mutex); |
| 794 | |
| 795 | if (watchdog_suspended) { |
| 796 | /* no parameter changes allowed while watchdog is suspended */ |
| 797 | err = -EAGAIN; |
| 798 | goto out; |
| 799 | } |
| 800 | |
| 801 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
| 802 | if (!err && write) { |
| 803 | /* Remove impossible cpus to keep sysctl output cleaner. */ |
| 804 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, |
| 805 | cpu_possible_mask); |
| 806 | |
| 807 | if (watchdog_running) { |
| 808 | /* |
| 809 | * Failure would be due to being unable to allocate |
| 810 | * a temporary cpumask, so we are likely not in a |
| 811 | * position to do much else to make things better. |
| 812 | */ |
| 813 | if (smpboot_update_cpumask_percpu_thread( |
| 814 | &watchdog_threads, &watchdog_cpumask) != 0) |
| 815 | pr_err("cpumask update failed\n"); |
| 816 | } |
| 817 | } |
| 818 | out: |
| 819 | mutex_unlock(&watchdog_proc_mutex); |
| 820 | put_online_cpus(); |
| 821 | return err; |
| 822 | } |
| 823 | |
| 824 | #endif /* CONFIG_SYSCTL */ |
| 825 | |
| 826 | void __init lockup_detector_init(void) |
| 827 | { |
| 828 | set_sample_period(); |
| 829 | |
| 830 | #ifdef CONFIG_NO_HZ_FULL |
| 831 | if (tick_nohz_full_enabled()) { |
| 832 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
| 833 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); |
| 834 | } else |
| 835 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); |
| 836 | #else |
| 837 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); |
| 838 | #endif |
| 839 | |
| 840 | if (watchdog_enabled) |
| 841 | watchdog_enable_all_cpus(); |
| 842 | } |