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) "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 /* Watchdog configuration */
  33 static DEFINE_MUTEX(watchdog_proc_mutex);
  34
  35 int __read_mostly nmi_watchdog_enabled;
  36
  37 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
  38 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED |
  39                                                 NMI_WATCHDOG_ENABLED;
  40 #else
  41 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
  42 #endif
  43
  44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
  45 /* boot commands */
  46 /*
  47  * Should we panic when a soft-lockup or hard-lockup occurs:
  48  */
  49 unsigned int __read_mostly hardlockup_panic =
  50                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  51 /*
  52  * We may not want to enable hard lockup detection by default in all cases,
  53  * for example when running the kernel as a guest on a hypervisor. In these
  54  * cases this function can be called to disable hard lockup detection. This
  55  * function should only be executed once by the boot processor before the
  56  * kernel command line parameters are parsed, because otherwise it is not
  57  * possible to override this in hardlockup_panic_setup().
  58  */
  59 void hardlockup_detector_disable(void)
  60 {
  61         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  62 }
  63
  64 static int __init hardlockup_panic_setup(char *str)
  65 {
  66         if (!strncmp(str, "panic", 5))
  67                 hardlockup_panic = 1;
  68         else if (!strncmp(str, "nopanic", 7))
  69                 hardlockup_panic = 0;
  70         else if (!strncmp(str, "0", 1))
  71                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
  72         else if (!strncmp(str, "1", 1))
  73                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
  74         return 1;
  75 }
  76 __setup("nmi_watchdog=", hardlockup_panic_setup);
  77
  78 #endif
  79
  80 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
  81 int __read_mostly soft_watchdog_enabled;
  82 #endif
  83
  84 int __read_mostly watchdog_user_enabled;
  85 int __read_mostly watchdog_thresh = 10;
  86
  87 #ifdef CONFIG_SMP
  88 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
  89 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  90 #endif
  91 struct cpumask watchdog_cpumask __read_mostly;
  92 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  93
  94 /*
  95  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  96  * are registered/started and is set to 0 when the watchdog threads are
  97  * unregistered/stopped, so it is an indicator whether the threads exist.
  98  */
  99 static int __read_mostly watchdog_running;
 100 /*
 101  * If a subsystem has a need to deactivate the watchdog temporarily, it
 102  * can use the suspend/resume interface to achieve this. The content of
 103  * the 'watchdog_suspended' variable reflects this state. Existing threads
 104  * are parked/unparked by the lockup_detector_{suspend|resume} functions
 105  * (see comment blocks pertaining to those functions for further details).
 106  *
 107  * 'watchdog_suspended' also prevents threads from being registered/started
 108  * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
 109  * of 'watchdog_running' cannot change while the watchdog is deactivated
 110  * temporarily (see related code in 'proc' handlers).
 111  */
 112 int __read_mostly watchdog_suspended;
 113
 114 /*
 115  * These functions can be overridden if an architecture implements its
 116  * own hardlockup detector.
 117  *
 118  * watchdog_nmi_enable/disable can be implemented to start and stop when
 119  * softlockup watchdog threads start and stop. The arch must select the
 120  * SOFTLOCKUP_DETECTOR Kconfig.
 121  */
 122 int __weak watchdog_nmi_enable(unsigned int cpu)
 123 {
 124         return 0;
 125 }
 126 void __weak watchdog_nmi_disable(unsigned int cpu)
 127 {
 128 }
 129
 130 /*
 131  * watchdog_nmi_reconfigure can be implemented to be notified after any
 132  * watchdog configuration change. The arch hardlockup watchdog should
 133  * respond to the following variables:
 134  * - nmi_watchdog_enabled
 135  * - watchdog_thresh
 136  * - watchdog_cpumask
 137  * - sysctl_hardlockup_all_cpu_backtrace
 138  * - hardlockup_panic
 139  * - watchdog_suspended
 140  */
 141 void __weak watchdog_nmi_reconfigure(void)
 142 {
 143 }
 144
 145
 146 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 147
 148 /* Helper for online, unparked cpus. */
 149 #define for_each_watchdog_cpu(cpu) \
 150         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
 151
 152 atomic_t watchdog_park_in_progress = ATOMIC_INIT(0);
 153
 154 static u64 __read_mostly sample_period;
 155
 156 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 157 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
 158 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 159 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 160 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
 161 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 162 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
 163 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
 164 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
 165 static unsigned long soft_lockup_nmi_warn;
 166
 167 unsigned int __read_mostly softlockup_panic =
 168                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
 169
 170 static int __init softlockup_panic_setup(char *str)
 171 {
 172         softlockup_panic = simple_strtoul(str, NULL, 0);
 173
 174         return 1;
 175 }
 176 __setup("softlockup_panic=", softlockup_panic_setup);
 177
 178 static int __init nowatchdog_setup(char *str)
 179 {
 180         watchdog_enabled = 0;
 181         return 1;
 182 }
 183 __setup("nowatchdog", nowatchdog_setup);
 184
 185 static int __init nosoftlockup_setup(char *str)
 186 {
 187         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
 188         return 1;
 189 }
 190 __setup("nosoftlockup", nosoftlockup_setup);
 191
 192 #ifdef CONFIG_SMP
 193 static int __init softlockup_all_cpu_backtrace_setup(char *str)
 194 {
 195         sysctl_softlockup_all_cpu_backtrace =
 196                 !!simple_strtol(str, NULL, 0);
 197         return 1;
 198 }
 199 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
 200 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 201 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
 202 {
 203         sysctl_hardlockup_all_cpu_backtrace =
 204                 !!simple_strtol(str, NULL, 0);
 205         return 1;
 206 }
 207 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
 208 #endif
 209 #endif
 210
 211 /*
 212  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 213  * lockups can have false positives under extreme conditions. So we generally
 214  * want a higher threshold for soft lockups than for hard lockups. So we couple
 215  * the thresholds with a factor: we make the soft threshold twice the amount of
 216  * time the hard threshold is.
 217  */
 218 static int get_softlockup_thresh(void)
 219 {
 220         return watchdog_thresh * 2;
 221 }
 222
 223 /*
 224  * Returns seconds, approximately.  We don't need nanosecond
 225  * resolution, and we don't need to waste time with a big divide when
 226  * 2^30ns == 1.074s.
 227  */
 228 static unsigned long get_timestamp(void)
 229 {
 230         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 231 }
 232
 233 static void set_sample_period(void)
 234 {
 235         /*
 236          * convert watchdog_thresh from seconds to ns
 237          * the divide by 5 is to give hrtimer several chances (two
 238          * or three with the current relation between the soft
 239          * and hard thresholds) to increment before the
 240          * hardlockup detector generates a warning
 241          */
 242         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 243         watchdog_update_hrtimer_threshold(sample_period);
 244 }
 245
 246 /* Commands for resetting the watchdog */
 247 static void __touch_watchdog(void)
 248 {
 249         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 250 }
 251
 252 /**
 253  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 254  *
 255  * Call when the scheduler may have stalled for legitimate reasons
 256  * preventing the watchdog task from executing - e.g. the scheduler
 257  * entering idle state.  This should only be used for scheduler events.
 258  * Use touch_softlockup_watchdog() for everything else.
 259  */
 260 void touch_softlockup_watchdog_sched(void)
 261 {
 262         /*
 263          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 264          * gets zeroed here, so use the raw_ operation.
 265          */
 266         raw_cpu_write(watchdog_touch_ts, 0);
 267 }
 268
 269 void touch_softlockup_watchdog(void)
 270 {
 271         touch_softlockup_watchdog_sched();
 272         wq_watchdog_touch(raw_smp_processor_id());
 273 }
 274 EXPORT_SYMBOL(touch_softlockup_watchdog);
 275
 276 void touch_all_softlockup_watchdogs(void)
 277 {
 278         int cpu;
 279
 280         /*
 281          * this is done lockless
 282          * do we care if a 0 races with a timestamp?
 283          * all it means is the softlock check starts one cycle later
 284          */
 285         for_each_watchdog_cpu(cpu)
 286                 per_cpu(watchdog_touch_ts, cpu) = 0;
 287         wq_watchdog_touch(-1);
 288 }
 289
 290 void touch_softlockup_watchdog_sync(void)
 291 {
 292         __this_cpu_write(softlockup_touch_sync, true);
 293         __this_cpu_write(watchdog_touch_ts, 0);
 294 }
 295
 296 static int is_softlockup(unsigned long touch_ts)
 297 {
 298         unsigned long now = get_timestamp();
 299
 300         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 301                 /* Warn about unreasonable delays. */
 302                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 303                         return now - touch_ts;
 304         }
 305         return 0;
 306 }
 307
 308 /* watchdog detector functions */
 309 bool is_hardlockup(void)
 310 {
 311         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 312
 313         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 314                 return true;
 315
 316         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 317         return false;
 318 }
 319
 320 static void watchdog_interrupt_count(void)
 321 {
 322         __this_cpu_inc(hrtimer_interrupts);
 323 }
 324
 325 static int watchdog_enable_all_cpus(void);
 326 static void watchdog_disable_all_cpus(void);
 327
 328 /* watchdog kicker functions */
 329 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 330 {
 331         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 332         struct pt_regs *regs = get_irq_regs();
 333         int duration;
 334         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 335
 336         if (atomic_read(&watchdog_park_in_progress) != 0)
 337                 return HRTIMER_NORESTART;
 338
 339         /* kick the hardlockup detector */
 340         watchdog_interrupt_count();
 341
 342         /* kick the softlockup detector */
 343         wake_up_process(__this_cpu_read(softlockup_watchdog));
 344
 345         /* .. and repeat */
 346         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 347
 348         if (touch_ts == 0) {
 349                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 350                         /*
 351                          * If the time stamp was touched atomically
 352                          * make sure the scheduler tick is up to date.
 353                          */
 354                         __this_cpu_write(softlockup_touch_sync, false);
 355                         sched_clock_tick();
 356                 }
 357
 358                 /* Clear the guest paused flag on watchdog reset */
 359                 kvm_check_and_clear_guest_paused();
 360                 __touch_watchdog();
 361                 return HRTIMER_RESTART;
 362         }
 363
 364         /* check for a softlockup
 365          * This is done by making sure a high priority task is
 366          * being scheduled.  The task touches the watchdog to
 367          * indicate it is getting cpu time.  If it hasn't then
 368          * this is a good indication some task is hogging the cpu
 369          */
 370         duration = is_softlockup(touch_ts);
 371         if (unlikely(duration)) {
 372                 /*
 373                  * If a virtual machine is stopped by the host it can look to
 374                  * the watchdog like a soft lockup, check to see if the host
 375                  * stopped the vm before we issue the warning
 376                  */
 377                 if (kvm_check_and_clear_guest_paused())
 378                         return HRTIMER_RESTART;
 379
 380                 /* only warn once */
 381                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 382                         /*
 383                          * When multiple processes are causing softlockups the
 384                          * softlockup detector only warns on the first one
 385                          * because the code relies on a full quiet cycle to
 386                          * re-arm.  The second process prevents the quiet cycle
 387                          * and never gets reported.  Use task pointers to detect
 388                          * this.
 389                          */
 390                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 391                             current) {
 392                                 __this_cpu_write(soft_watchdog_warn, false);
 393                                 __touch_watchdog();
 394                         }
 395                         return HRTIMER_RESTART;
 396                 }
 397
 398                 if (softlockup_all_cpu_backtrace) {
 399                         /* Prevent multiple soft-lockup reports if one cpu is already
 400                          * engaged in dumping cpu back traces
 401                          */
 402                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 403                                 /* Someone else will report us. Let's give up */
 404                                 __this_cpu_write(soft_watchdog_warn, true);
 405                                 return HRTIMER_RESTART;
 406                         }
 407                 }
 408
 409                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 410                         smp_processor_id(), duration,
 411                         current->comm, task_pid_nr(current));
 412                 __this_cpu_write(softlockup_task_ptr_saved, current);
 413                 print_modules();
 414                 print_irqtrace_events(current);
 415                 if (regs)
 416                         show_regs(regs);
 417                 else
 418                         dump_stack();
 419
 420                 if (softlockup_all_cpu_backtrace) {
 421                         /* Avoid generating two back traces for current
 422                          * given that one is already made above
 423                          */
 424                         trigger_allbutself_cpu_backtrace();
 425
 426                         clear_bit(0, &soft_lockup_nmi_warn);
 427                         /* Barrier to sync with other cpus */
 428                         smp_mb__after_atomic();
 429                 }
 430
 431                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 432                 if (softlockup_panic)
 433                         panic("softlockup: hung tasks");
 434                 __this_cpu_write(soft_watchdog_warn, true);
 435         } else
 436                 __this_cpu_write(soft_watchdog_warn, false);
 437
 438         return HRTIMER_RESTART;
 439 }
 440
 441 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 442 {
 443         struct sched_param param = { .sched_priority = prio };
 444
 445         sched_setscheduler(current, policy, &param);
 446 }
 447
 448 static void watchdog_enable(unsigned int cpu)
 449 {
 450         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 451
 452         /* kick off the timer for the hardlockup detector */
 453         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 454         hrtimer->function = watchdog_timer_fn;
 455
 456         /* Enable the perf event */
 457         watchdog_nmi_enable(cpu);
 458
 459         /* done here because hrtimer_start can only pin to smp_processor_id() */
 460         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 461                       HRTIMER_MODE_REL_PINNED);
 462
 463         /* initialize timestamp */
 464         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 465         __touch_watchdog();
 466 }
 467
 468 static void watchdog_disable(unsigned int cpu)
 469 {
 470         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 471
 472         watchdog_set_prio(SCHED_NORMAL, 0);
 473         hrtimer_cancel(hrtimer);
 474         /* disable the perf event */
 475         watchdog_nmi_disable(cpu);
 476 }
 477
 478 static void watchdog_cleanup(unsigned int cpu, bool online)
 479 {
 480         watchdog_disable(cpu);
 481 }
 482
 483 static int watchdog_should_run(unsigned int cpu)
 484 {
 485         return __this_cpu_read(hrtimer_interrupts) !=
 486                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 487 }
 488
 489 /*
 490  * The watchdog thread function - touches the timestamp.
 491  *
 492  * It only runs once every sample_period seconds (4 seconds by
 493  * default) to reset the softlockup timestamp. If this gets delayed
 494  * for more than 2*watchdog_thresh seconds then the debug-printout
 495  * triggers in watchdog_timer_fn().
 496  */
 497 static void watchdog(unsigned int cpu)
 498 {
 499         __this_cpu_write(soft_lockup_hrtimer_cnt,
 500                          __this_cpu_read(hrtimer_interrupts));
 501         __touch_watchdog();
 502
 503         /*
 504          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 505          * failure path. Check for failures that can occur asynchronously -
 506          * for example, when CPUs are on-lined - and shut down the hardware
 507          * perf event on each CPU accordingly.
 508          *
 509          * The only non-obvious place this bit can be cleared is through
 510          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 511          * pr_info here would be too noisy as it would result in a message
 512          * every few seconds if the hardlockup was disabled but the softlockup
 513          * enabled.
 514          */
 515         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 516                 watchdog_nmi_disable(cpu);
 517 }
 518
 519 static struct smp_hotplug_thread watchdog_threads = {
 520         .store                  = &softlockup_watchdog,
 521         .thread_should_run      = watchdog_should_run,
 522         .thread_fn              = watchdog,
 523         .thread_comm            = "watchdog/%u",
 524         .setup                  = watchdog_enable,
 525         .cleanup                = watchdog_cleanup,
 526         .park                   = watchdog_disable,
 527         .unpark                 = watchdog_enable,
 528 };
 529
 530 /*
 531  * park all watchdog threads that are specified in 'watchdog_cpumask'
 532  *
 533  * This function returns an error if kthread_park() of a watchdog thread
 534  * fails. In this situation, the watchdog threads of some CPUs can already
 535  * be parked and the watchdog threads of other CPUs can still be runnable.
 536  * Callers are expected to handle this special condition as appropriate in
 537  * their context.
 538  *
 539  * This function may only be called in a context that is protected against
 540  * races with CPU hotplug - for example, via get_online_cpus().
 541  */
 542 static int watchdog_park_threads(void)
 543 {
 544         int cpu, ret = 0;
 545
 546         atomic_set(&watchdog_park_in_progress, 1);
 547
 548         for_each_watchdog_cpu(cpu) {
 549                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 550                 if (ret)
 551                         break;
 552         }
 553
 554         atomic_set(&watchdog_park_in_progress, 0);
 555
 556         return ret;
 557 }
 558
 559 /*
 560  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 561  *
 562  * This function may only be called in a context that is protected against
 563  * races with CPU hotplug - for example, via get_online_cpus().
 564  */
 565 static void watchdog_unpark_threads(void)
 566 {
 567         int cpu;
 568
 569         for_each_watchdog_cpu(cpu)
 570                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 571 }
 572
 573 static int update_watchdog_all_cpus(void)
 574 {
 575         int ret;
 576
 577         ret = watchdog_park_threads();
 578         if (ret)
 579                 return ret;
 580
 581         watchdog_unpark_threads();
 582
 583         return 0;
 584 }
 585
 586 static int watchdog_enable_all_cpus(void)
 587 {
 588         int err = 0;
 589
 590         if (!watchdog_running) {
 591                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 592                                                              &watchdog_cpumask);
 593                 if (err)
 594                         pr_err("Failed to create watchdog threads, disabled\n");
 595                 else
 596                         watchdog_running = 1;
 597         } else {
 598                 /*
 599                  * Enable/disable the lockup detectors or
 600                  * change the sample period 'on the fly'.
 601                  */
 602                 err = update_watchdog_all_cpus();
 603
 604                 if (err) {
 605                         watchdog_disable_all_cpus();
 606                         pr_err("Failed to update lockup detectors, disabled\n");
 607                 }
 608         }
 609
 610         if (err)
 611                 watchdog_enabled = 0;
 612
 613         return err;
 614 }
 615
 616 static void watchdog_disable_all_cpus(void)
 617 {
 618         if (watchdog_running) {
 619                 watchdog_running = 0;
 620                 smpboot_unregister_percpu_thread(&watchdog_threads);
 621         }
 622 }
 623
 624 #ifdef CONFIG_SYSCTL
 625 static int watchdog_update_cpus(void)
 626 {
 627         return smpboot_update_cpumask_percpu_thread(
 628                     &watchdog_threads, &watchdog_cpumask);
 629 }
 630 #endif
 631
 632 #else /* SOFTLOCKUP */
 633 static int watchdog_park_threads(void)
 634 {
 635         return 0;
 636 }
 637
 638 static void watchdog_unpark_threads(void)
 639 {
 640 }
 641
 642 static int watchdog_enable_all_cpus(void)
 643 {
 644         return 0;
 645 }
 646
 647 static void watchdog_disable_all_cpus(void)
 648 {
 649 }
 650
 651 #ifdef CONFIG_SYSCTL
 652 static int watchdog_update_cpus(void)
 653 {
 654         return 0;
 655 }
 656 #endif
 657
 658 static void set_sample_period(void)
 659 {
 660 }
 661 #endif /* SOFTLOCKUP */
 662
 663 /*
 664  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 665  */
 666 int lockup_detector_suspend(void)
 667 {
 668         int ret = 0;
 669
 670         get_online_cpus();
 671         mutex_lock(&watchdog_proc_mutex);
 672         /*
 673          * Multiple suspend requests can be active in parallel (counted by
 674          * the 'watchdog_suspended' variable). If the watchdog threads are
 675          * running, the first caller takes care that they will be parked.
 676          * The state of 'watchdog_running' cannot change while a suspend
 677          * request is active (see related code in 'proc' handlers).
 678          */
 679         if (watchdog_running && !watchdog_suspended)
 680                 ret = watchdog_park_threads();
 681
 682         if (ret == 0)
 683                 watchdog_suspended++;
 684         else {
 685                 watchdog_disable_all_cpus();
 686                 pr_err("Failed to suspend lockup detectors, disabled\n");
 687                 watchdog_enabled = 0;
 688         }
 689
 690         watchdog_nmi_reconfigure();
 691
 692         mutex_unlock(&watchdog_proc_mutex);
 693
 694         return ret;
 695 }
 696
 697 /*
 698  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 699  */
 700 void lockup_detector_resume(void)
 701 {
 702         mutex_lock(&watchdog_proc_mutex);
 703
 704         watchdog_suspended--;
 705         /*
 706          * The watchdog threads are unparked if they were previously running
 707          * and if there is no more active suspend request.
 708          */
 709         if (watchdog_running && !watchdog_suspended)
 710                 watchdog_unpark_threads();
 711
 712         watchdog_nmi_reconfigure();
 713
 714         mutex_unlock(&watchdog_proc_mutex);
 715         put_online_cpus();
 716 }
 717
 718 #ifdef CONFIG_SYSCTL
 719
 720 /*
 721  * Update the run state of the lockup detectors.
 722  */
 723 static int proc_watchdog_update(void)
 724 {
 725         int err = 0;
 726
 727         /*
 728          * Watchdog threads won't be started if they are already active.
 729          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 730          * care of this. If those threads are already active, the sample
 731          * period will be updated and the lockup detectors will be enabled
 732          * or disabled 'on the fly'.
 733          */
 734         if (watchdog_enabled && watchdog_thresh)
 735                 err = watchdog_enable_all_cpus();
 736         else
 737                 watchdog_disable_all_cpus();
 738
 739         watchdog_nmi_reconfigure();
 740
 741         return err;
 742
 743 }
 744
 745 /*
 746  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 747  *
 748  * caller             | table->data points to | 'which' contains the flag(s)
 749  * -------------------|-----------------------|-----------------------------
 750  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 751  *                    |                       | with SOFT_WATCHDOG_ENABLED
 752  * -------------------|-----------------------|-----------------------------
 753  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 754  * -------------------|-----------------------|-----------------------------
 755  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 756  */
 757 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 758                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 759 {
 760         int err, old, new;
 761         int *watchdog_param = (int *)table->data;
 762
 763         get_online_cpus();
 764         mutex_lock(&watchdog_proc_mutex);
 765
 766         if (watchdog_suspended) {
 767                 /* no parameter changes allowed while watchdog is suspended */
 768                 err = -EAGAIN;
 769                 goto out;
 770         }
 771
 772         /*
 773          * If the parameter is being read return the state of the corresponding
 774          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 775          * run state of the lockup detectors.
 776          */
 777         if (!write) {
 778                 *watchdog_param = (watchdog_enabled & which) != 0;
 779                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 780         } else {
 781                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 782                 if (err)
 783                         goto out;
 784
 785                 /*
 786                  * There is a race window between fetching the current value
 787                  * from 'watchdog_enabled' and storing the new value. During
 788                  * this race window, watchdog_nmi_enable() can sneak in and
 789                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 790                  * The 'cmpxchg' detects this race and the loop retries.
 791                  */
 792                 do {
 793                         old = watchdog_enabled;
 794                         /*
 795                          * If the parameter value is not zero set the
 796                          * corresponding bit(s), else clear it(them).
 797                          */
 798                         if (*watchdog_param)
 799                                 new = old | which;
 800                         else
 801                                 new = old & ~which;
 802                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 803
 804                 /*
 805                  * Update the run state of the lockup detectors. There is _no_
 806                  * need to check the value returned by proc_watchdog_update()
 807                  * and to restore the previous value of 'watchdog_enabled' as
 808                  * both lockup detectors are disabled if proc_watchdog_update()
 809                  * returns an error.
 810                  */
 811                 if (old == new)
 812                         goto out;
 813
 814                 err = proc_watchdog_update();
 815         }
 816 out:
 817         mutex_unlock(&watchdog_proc_mutex);
 818         put_online_cpus();
 819         return err;
 820 }
 821
 822 /*
 823  * /proc/sys/kernel/watchdog
 824  */
 825 int proc_watchdog(struct ctl_table *table, int write,
 826                   void __user *buffer, size_t *lenp, loff_t *ppos)
 827 {
 828         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 829                                     table, write, buffer, lenp, ppos);
 830 }
 831
 832 /*
 833  * /proc/sys/kernel/nmi_watchdog
 834  */
 835 int proc_nmi_watchdog(struct ctl_table *table, int write,
 836                       void __user *buffer, size_t *lenp, loff_t *ppos)
 837 {
 838         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 839                                     table, write, buffer, lenp, ppos);
 840 }
 841
 842 /*
 843  * /proc/sys/kernel/soft_watchdog
 844  */
 845 int proc_soft_watchdog(struct ctl_table *table, int write,
 846                         void __user *buffer, size_t *lenp, loff_t *ppos)
 847 {
 848         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 849                                     table, write, buffer, lenp, ppos);
 850 }
 851
 852 /*
 853  * /proc/sys/kernel/watchdog_thresh
 854  */
 855 int proc_watchdog_thresh(struct ctl_table *table, int write,
 856                          void __user *buffer, size_t *lenp, loff_t *ppos)
 857 {
 858         int err, old, new;
 859
 860         get_online_cpus();
 861         mutex_lock(&watchdog_proc_mutex);
 862
 863         if (watchdog_suspended) {
 864                 /* no parameter changes allowed while watchdog is suspended */
 865                 err = -EAGAIN;
 866                 goto out;
 867         }
 868
 869         old = ACCESS_ONCE(watchdog_thresh);
 870         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 871
 872         if (err || !write)
 873                 goto out;
 874
 875         /*
 876          * Update the sample period. Restore on failure.
 877          */
 878         new = ACCESS_ONCE(watchdog_thresh);
 879         if (old == new)
 880                 goto out;
 881
 882         set_sample_period();
 883         err = proc_watchdog_update();
 884         if (err) {
 885                 watchdog_thresh = old;
 886                 set_sample_period();
 887         }
 888 out:
 889         mutex_unlock(&watchdog_proc_mutex);
 890         put_online_cpus();
 891         return err;
 892 }
 893
 894 /*
 895  * The cpumask is the mask of possible cpus that the watchdog can run
 896  * on, not the mask of cpus it is actually running on.  This allows the
 897  * user to specify a mask that will include cpus that have not yet
 898  * been brought online, if desired.
 899  */
 900 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 901                           void __user *buffer, size_t *lenp, loff_t *ppos)
 902 {
 903         int err;
 904
 905         get_online_cpus();
 906         mutex_lock(&watchdog_proc_mutex);
 907
 908         if (watchdog_suspended) {
 909                 /* no parameter changes allowed while watchdog is suspended */
 910                 err = -EAGAIN;
 911                 goto out;
 912         }
 913
 914         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 915         if (!err && write) {
 916                 /* Remove impossible cpus to keep sysctl output cleaner. */
 917                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
 918                             cpu_possible_mask);
 919
 920                 if (watchdog_running) {
 921                         /*
 922                          * Failure would be due to being unable to allocate
 923                          * a temporary cpumask, so we are likely not in a
 924                          * position to do much else to make things better.
 925                          */
 926                         if (watchdog_update_cpus() != 0)
 927                                 pr_err("cpumask update failed\n");
 928                 }
 929
 930                 watchdog_nmi_reconfigure();
 931         }
 932 out:
 933         mutex_unlock(&watchdog_proc_mutex);
 934         put_online_cpus();
 935         return err;
 936 }
 937
 938 #endif /* CONFIG_SYSCTL */
 939
 940 void __init lockup_detector_init(void)
 941 {
 942         set_sample_period();
 943
 944 #ifdef CONFIG_NO_HZ_FULL
 945         if (tick_nohz_full_enabled()) {
 946                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 947                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
 948         } else
 949                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 950 #else
 951         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 952 #endif
 953
 954         if (watchdog_enabled)
 955                 watchdog_enable_all_cpus();
 956 }