kernel/watchdog.c

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