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 }
 244
 245 /* Commands for resetting the watchdog */
 246 static void __touch_watchdog(void)
 247 {
 248         __this_cpu_write(watchdog_touch_ts, get_timestamp());
 249 }
 250
 251 /**
 252  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 253  *
 254  * Call when the scheduler may have stalled for legitimate reasons
 255  * preventing the watchdog task from executing - e.g. the scheduler
 256  * entering idle state.  This should only be used for scheduler events.
 257  * Use touch_softlockup_watchdog() for everything else.
 258  */
 259 void touch_softlockup_watchdog_sched(void)
 260 {
 261         /*
 262          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
 263          * gets zeroed here, so use the raw_ operation.
 264          */
 265         raw_cpu_write(watchdog_touch_ts, 0);
 266 }
 267
 268 void touch_softlockup_watchdog(void)
 269 {
 270         touch_softlockup_watchdog_sched();
 271         wq_watchdog_touch(raw_smp_processor_id());
 272 }
 273 EXPORT_SYMBOL(touch_softlockup_watchdog);
 274
 275 void touch_all_softlockup_watchdogs(void)
 276 {
 277         int cpu;
 278
 279         /*
 280          * this is done lockless
 281          * do we care if a 0 races with a timestamp?
 282          * all it means is the softlock check starts one cycle later
 283          */
 284         for_each_watchdog_cpu(cpu)
 285                 per_cpu(watchdog_touch_ts, cpu) = 0;
 286         wq_watchdog_touch(-1);
 287 }
 288
 289 void touch_softlockup_watchdog_sync(void)
 290 {
 291         __this_cpu_write(softlockup_touch_sync, true);
 292         __this_cpu_write(watchdog_touch_ts, 0);
 293 }
 294
 295 static int is_softlockup(unsigned long touch_ts)
 296 {
 297         unsigned long now = get_timestamp();
 298
 299         if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
 300                 /* Warn about unreasonable delays. */
 301                 if (time_after(now, touch_ts + get_softlockup_thresh()))
 302                         return now - touch_ts;
 303         }
 304         return 0;
 305 }
 306
 307 /* watchdog detector functions */
 308 bool is_hardlockup(void)
 309 {
 310         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 311
 312         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 313                 return true;
 314
 315         __this_cpu_write(hrtimer_interrupts_saved, hrint);
 316         return false;
 317 }
 318
 319 static void watchdog_interrupt_count(void)
 320 {
 321         __this_cpu_inc(hrtimer_interrupts);
 322 }
 323
 324 static int watchdog_enable_all_cpus(void);
 325 static void watchdog_disable_all_cpus(void);
 326
 327 /* watchdog kicker functions */
 328 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 329 {
 330         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 331         struct pt_regs *regs = get_irq_regs();
 332         int duration;
 333         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 334
 335         if (atomic_read(&watchdog_park_in_progress) != 0)
 336                 return HRTIMER_NORESTART;
 337
 338         /* kick the hardlockup detector */
 339         watchdog_interrupt_count();
 340
 341         /* kick the softlockup detector */
 342         wake_up_process(__this_cpu_read(softlockup_watchdog));
 343
 344         /* .. and repeat */
 345         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 346
 347         if (touch_ts == 0) {
 348                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 349                         /*
 350                          * If the time stamp was touched atomically
 351                          * make sure the scheduler tick is up to date.
 352                          */
 353                         __this_cpu_write(softlockup_touch_sync, false);
 354                         sched_clock_tick();
 355                 }
 356
 357                 /* Clear the guest paused flag on watchdog reset */
 358                 kvm_check_and_clear_guest_paused();
 359                 __touch_watchdog();
 360                 return HRTIMER_RESTART;
 361         }
 362
 363         /* check for a softlockup
 364          * This is done by making sure a high priority task is
 365          * being scheduled.  The task touches the watchdog to
 366          * indicate it is getting cpu time.  If it hasn't then
 367          * this is a good indication some task is hogging the cpu
 368          */
 369         duration = is_softlockup(touch_ts);
 370         if (unlikely(duration)) {
 371                 /*
 372                  * If a virtual machine is stopped by the host it can look to
 373                  * the watchdog like a soft lockup, check to see if the host
 374                  * stopped the vm before we issue the warning
 375                  */
 376                 if (kvm_check_and_clear_guest_paused())
 377                         return HRTIMER_RESTART;
 378
 379                 /* only warn once */
 380                 if (__this_cpu_read(soft_watchdog_warn) == true) {
 381                         /*
 382                          * When multiple processes are causing softlockups the
 383                          * softlockup detector only warns on the first one
 384                          * because the code relies on a full quiet cycle to
 385                          * re-arm.  The second process prevents the quiet cycle
 386                          * and never gets reported.  Use task pointers to detect
 387                          * this.
 388                          */
 389                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
 390                             current) {
 391                                 __this_cpu_write(soft_watchdog_warn, false);
 392                                 __touch_watchdog();
 393                         }
 394                         return HRTIMER_RESTART;
 395                 }
 396
 397                 if (softlockup_all_cpu_backtrace) {
 398                         /* Prevent multiple soft-lockup reports if one cpu is already
 399                          * engaged in dumping cpu back traces
 400                          */
 401                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
 402                                 /* Someone else will report us. Let's give up */
 403                                 __this_cpu_write(soft_watchdog_warn, true);
 404                                 return HRTIMER_RESTART;
 405                         }
 406                 }
 407
 408                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 409                         smp_processor_id(), duration,
 410                         current->comm, task_pid_nr(current));
 411                 __this_cpu_write(softlockup_task_ptr_saved, current);
 412                 print_modules();
 413                 print_irqtrace_events(current);
 414                 if (regs)
 415                         show_regs(regs);
 416                 else
 417                         dump_stack();
 418
 419                 if (softlockup_all_cpu_backtrace) {
 420                         /* Avoid generating two back traces for current
 421                          * given that one is already made above
 422                          */
 423                         trigger_allbutself_cpu_backtrace();
 424
 425                         clear_bit(0, &soft_lockup_nmi_warn);
 426                         /* Barrier to sync with other cpus */
 427                         smp_mb__after_atomic();
 428                 }
 429
 430                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 431                 if (softlockup_panic)
 432                         panic("softlockup: hung tasks");
 433                 __this_cpu_write(soft_watchdog_warn, true);
 434         } else
 435                 __this_cpu_write(soft_watchdog_warn, false);
 436
 437         return HRTIMER_RESTART;
 438 }
 439
 440 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
 441 {
 442         struct sched_param param = { .sched_priority = prio };
 443
 444         sched_setscheduler(current, policy, &param);
 445 }
 446
 447 static void watchdog_enable(unsigned int cpu)
 448 {
 449         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 450
 451         /* kick off the timer for the hardlockup detector */
 452         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 453         hrtimer->function = watchdog_timer_fn;
 454
 455         /* Enable the perf event */
 456         watchdog_nmi_enable(cpu);
 457
 458         /* done here because hrtimer_start can only pin to smp_processor_id() */
 459         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 460                       HRTIMER_MODE_REL_PINNED);
 461
 462         /* initialize timestamp */
 463         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
 464         __touch_watchdog();
 465 }
 466
 467 static void watchdog_disable(unsigned int cpu)
 468 {
 469         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
 470
 471         watchdog_set_prio(SCHED_NORMAL, 0);
 472         hrtimer_cancel(hrtimer);
 473         /* disable the perf event */
 474         watchdog_nmi_disable(cpu);
 475 }
 476
 477 static void watchdog_cleanup(unsigned int cpu, bool online)
 478 {
 479         watchdog_disable(cpu);
 480 }
 481
 482 static int watchdog_should_run(unsigned int cpu)
 483 {
 484         return __this_cpu_read(hrtimer_interrupts) !=
 485                 __this_cpu_read(soft_lockup_hrtimer_cnt);
 486 }
 487
 488 /*
 489  * The watchdog thread function - touches the timestamp.
 490  *
 491  * It only runs once every sample_period seconds (4 seconds by
 492  * default) to reset the softlockup timestamp. If this gets delayed
 493  * for more than 2*watchdog_thresh seconds then the debug-printout
 494  * triggers in watchdog_timer_fn().
 495  */
 496 static void watchdog(unsigned int cpu)
 497 {
 498         __this_cpu_write(soft_lockup_hrtimer_cnt,
 499                          __this_cpu_read(hrtimer_interrupts));
 500         __touch_watchdog();
 501
 502         /*
 503          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
 504          * failure path. Check for failures that can occur asynchronously -
 505          * for example, when CPUs are on-lined - and shut down the hardware
 506          * perf event on each CPU accordingly.
 507          *
 508          * The only non-obvious place this bit can be cleared is through
 509          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
 510          * pr_info here would be too noisy as it would result in a message
 511          * every few seconds if the hardlockup was disabled but the softlockup
 512          * enabled.
 513          */
 514         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 515                 watchdog_nmi_disable(cpu);
 516 }
 517
 518 static struct smp_hotplug_thread watchdog_threads = {
 519         .store                  = &softlockup_watchdog,
 520         .thread_should_run      = watchdog_should_run,
 521         .thread_fn              = watchdog,
 522         .thread_comm            = "watchdog/%u",
 523         .setup                  = watchdog_enable,
 524         .cleanup                = watchdog_cleanup,
 525         .park                   = watchdog_disable,
 526         .unpark                 = watchdog_enable,
 527 };
 528
 529 /*
 530  * park all watchdog threads that are specified in 'watchdog_cpumask'
 531  *
 532  * This function returns an error if kthread_park() of a watchdog thread
 533  * fails. In this situation, the watchdog threads of some CPUs can already
 534  * be parked and the watchdog threads of other CPUs can still be runnable.
 535  * Callers are expected to handle this special condition as appropriate in
 536  * their context.
 537  *
 538  * This function may only be called in a context that is protected against
 539  * races with CPU hotplug - for example, via get_online_cpus().
 540  */
 541 static int watchdog_park_threads(void)
 542 {
 543         int cpu, ret = 0;
 544
 545         atomic_set(&watchdog_park_in_progress, 1);
 546
 547         for_each_watchdog_cpu(cpu) {
 548                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 549                 if (ret)
 550                         break;
 551         }
 552
 553         atomic_set(&watchdog_park_in_progress, 0);
 554
 555         return ret;
 556 }
 557
 558 /*
 559  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
 560  *
 561  * This function may only be called in a context that is protected against
 562  * races with CPU hotplug - for example, via get_online_cpus().
 563  */
 564 static void watchdog_unpark_threads(void)
 565 {
 566         int cpu;
 567
 568         for_each_watchdog_cpu(cpu)
 569                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
 570 }
 571
 572 static int update_watchdog_all_cpus(void)
 573 {
 574         int ret;
 575
 576         ret = watchdog_park_threads();
 577         if (ret)
 578                 return ret;
 579
 580         watchdog_unpark_threads();
 581
 582         return 0;
 583 }
 584
 585 static int watchdog_enable_all_cpus(void)
 586 {
 587         int err = 0;
 588
 589         if (!watchdog_running) {
 590                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
 591                                                              &watchdog_cpumask);
 592                 if (err)
 593                         pr_err("Failed to create watchdog threads, disabled\n");
 594                 else
 595                         watchdog_running = 1;
 596         } else {
 597                 /*
 598                  * Enable/disable the lockup detectors or
 599                  * change the sample period 'on the fly'.
 600                  */
 601                 err = update_watchdog_all_cpus();
 602
 603                 if (err) {
 604                         watchdog_disable_all_cpus();
 605                         pr_err("Failed to update lockup detectors, disabled\n");
 606                 }
 607         }
 608
 609         if (err)
 610                 watchdog_enabled = 0;
 611
 612         return err;
 613 }
 614
 615 static void watchdog_disable_all_cpus(void)
 616 {
 617         if (watchdog_running) {
 618                 watchdog_running = 0;
 619                 smpboot_unregister_percpu_thread(&watchdog_threads);
 620         }
 621 }
 622
 623 #ifdef CONFIG_SYSCTL
 624 static int watchdog_update_cpus(void)
 625 {
 626         return smpboot_update_cpumask_percpu_thread(
 627                     &watchdog_threads, &watchdog_cpumask);
 628 }
 629 #endif
 630
 631 #else /* SOFTLOCKUP */
 632 static int watchdog_park_threads(void)
 633 {
 634         return 0;
 635 }
 636
 637 static void watchdog_unpark_threads(void)
 638 {
 639 }
 640
 641 static int watchdog_enable_all_cpus(void)
 642 {
 643         return 0;
 644 }
 645
 646 static void watchdog_disable_all_cpus(void)
 647 {
 648 }
 649
 650 #ifdef CONFIG_SYSCTL
 651 static int watchdog_update_cpus(void)
 652 {
 653         return 0;
 654 }
 655 #endif
 656
 657 static void set_sample_period(void)
 658 {
 659 }
 660 #endif /* SOFTLOCKUP */
 661
 662 /*
 663  * Suspend the hard and soft lockup detector by parking the watchdog threads.
 664  */
 665 int lockup_detector_suspend(void)
 666 {
 667         int ret = 0;
 668
 669         get_online_cpus();
 670         mutex_lock(&watchdog_proc_mutex);
 671         /*
 672          * Multiple suspend requests can be active in parallel (counted by
 673          * the 'watchdog_suspended' variable). If the watchdog threads are
 674          * running, the first caller takes care that they will be parked.
 675          * The state of 'watchdog_running' cannot change while a suspend
 676          * request is active (see related code in 'proc' handlers).
 677          */
 678         if (watchdog_running && !watchdog_suspended)
 679                 ret = watchdog_park_threads();
 680
 681         if (ret == 0)
 682                 watchdog_suspended++;
 683         else {
 684                 watchdog_disable_all_cpus();
 685                 pr_err("Failed to suspend lockup detectors, disabled\n");
 686                 watchdog_enabled = 0;
 687         }
 688
 689         watchdog_nmi_reconfigure();
 690
 691         mutex_unlock(&watchdog_proc_mutex);
 692
 693         return ret;
 694 }
 695
 696 /*
 697  * Resume the hard and soft lockup detector by unparking the watchdog threads.
 698  */
 699 void lockup_detector_resume(void)
 700 {
 701         mutex_lock(&watchdog_proc_mutex);
 702
 703         watchdog_suspended--;
 704         /*
 705          * The watchdog threads are unparked if they were previously running
 706          * and if there is no more active suspend request.
 707          */
 708         if (watchdog_running && !watchdog_suspended)
 709                 watchdog_unpark_threads();
 710
 711         watchdog_nmi_reconfigure();
 712
 713         mutex_unlock(&watchdog_proc_mutex);
 714         put_online_cpus();
 715 }
 716
 717 #ifdef CONFIG_SYSCTL
 718
 719 /*
 720  * Update the run state of the lockup detectors.
 721  */
 722 static int proc_watchdog_update(void)
 723 {
 724         int err = 0;
 725
 726         /*
 727          * Watchdog threads won't be started if they are already active.
 728          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
 729          * care of this. If those threads are already active, the sample
 730          * period will be updated and the lockup detectors will be enabled
 731          * or disabled 'on the fly'.
 732          */
 733         if (watchdog_enabled && watchdog_thresh)
 734                 err = watchdog_enable_all_cpus();
 735         else
 736                 watchdog_disable_all_cpus();
 737
 738         watchdog_nmi_reconfigure();
 739
 740         return err;
 741
 742 }
 743
 744 /*
 745  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 746  *
 747  * caller             | table->data points to | 'which' contains the flag(s)
 748  * -------------------|-----------------------|-----------------------------
 749  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
 750  *                    |                       | with SOFT_WATCHDOG_ENABLED
 751  * -------------------|-----------------------|-----------------------------
 752  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
 753  * -------------------|-----------------------|-----------------------------
 754  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
 755  */
 756 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
 757                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 758 {
 759         int err, old, new;
 760         int *watchdog_param = (int *)table->data;
 761
 762         get_online_cpus();
 763         mutex_lock(&watchdog_proc_mutex);
 764
 765         if (watchdog_suspended) {
 766                 /* no parameter changes allowed while watchdog is suspended */
 767                 err = -EAGAIN;
 768                 goto out;
 769         }
 770
 771         /*
 772          * If the parameter is being read return the state of the corresponding
 773          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
 774          * run state of the lockup detectors.
 775          */
 776         if (!write) {
 777                 *watchdog_param = (watchdog_enabled & which) != 0;
 778                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 779         } else {
 780                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 781                 if (err)
 782                         goto out;
 783
 784                 /*
 785                  * There is a race window between fetching the current value
 786                  * from 'watchdog_enabled' and storing the new value. During
 787                  * this race window, watchdog_nmi_enable() can sneak in and
 788                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
 789                  * The 'cmpxchg' detects this race and the loop retries.
 790                  */
 791                 do {
 792                         old = watchdog_enabled;
 793                         /*
 794                          * If the parameter value is not zero set the
 795                          * corresponding bit(s), else clear it(them).
 796                          */
 797                         if (*watchdog_param)
 798                                 new = old | which;
 799                         else
 800                                 new = old & ~which;
 801                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
 802
 803                 /*
 804                  * Update the run state of the lockup detectors. There is _no_
 805                  * need to check the value returned by proc_watchdog_update()
 806                  * and to restore the previous value of 'watchdog_enabled' as
 807                  * both lockup detectors are disabled if proc_watchdog_update()
 808                  * returns an error.
 809                  */
 810                 if (old == new)
 811                         goto out;
 812
 813                 err = proc_watchdog_update();
 814         }
 815 out:
 816         mutex_unlock(&watchdog_proc_mutex);
 817         put_online_cpus();
 818         return err;
 819 }
 820
 821 /*
 822  * /proc/sys/kernel/watchdog
 823  */
 824 int proc_watchdog(struct ctl_table *table, int write,
 825                   void __user *buffer, size_t *lenp, loff_t *ppos)
 826 {
 827         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
 828                                     table, write, buffer, lenp, ppos);
 829 }
 830
 831 /*
 832  * /proc/sys/kernel/nmi_watchdog
 833  */
 834 int proc_nmi_watchdog(struct ctl_table *table, int write,
 835                       void __user *buffer, size_t *lenp, loff_t *ppos)
 836 {
 837         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
 838                                     table, write, buffer, lenp, ppos);
 839 }
 840
 841 /*
 842  * /proc/sys/kernel/soft_watchdog
 843  */
 844 int proc_soft_watchdog(struct ctl_table *table, int write,
 845                         void __user *buffer, size_t *lenp, loff_t *ppos)
 846 {
 847         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
 848                                     table, write, buffer, lenp, ppos);
 849 }
 850
 851 /*
 852  * /proc/sys/kernel/watchdog_thresh
 853  */
 854 int proc_watchdog_thresh(struct ctl_table *table, int write,
 855                          void __user *buffer, size_t *lenp, loff_t *ppos)
 856 {
 857         int err, old, new;
 858
 859         get_online_cpus();
 860         mutex_lock(&watchdog_proc_mutex);
 861
 862         if (watchdog_suspended) {
 863                 /* no parameter changes allowed while watchdog is suspended */
 864                 err = -EAGAIN;
 865                 goto out;
 866         }
 867
 868         old = ACCESS_ONCE(watchdog_thresh);
 869         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 870
 871         if (err || !write)
 872                 goto out;
 873
 874         /*
 875          * Update the sample period. Restore on failure.
 876          */
 877         new = ACCESS_ONCE(watchdog_thresh);
 878         if (old == new)
 879                 goto out;
 880
 881         set_sample_period();
 882         err = proc_watchdog_update();
 883         if (err) {
 884                 watchdog_thresh = old;
 885                 set_sample_period();
 886         }
 887 out:
 888         mutex_unlock(&watchdog_proc_mutex);
 889         put_online_cpus();
 890         return err;
 891 }
 892
 893 /*
 894  * The cpumask is the mask of possible cpus that the watchdog can run
 895  * on, not the mask of cpus it is actually running on.  This allows the
 896  * user to specify a mask that will include cpus that have not yet
 897  * been brought online, if desired.
 898  */
 899 int proc_watchdog_cpumask(struct ctl_table *table, int write,
 900                           void __user *buffer, size_t *lenp, loff_t *ppos)
 901 {
 902         int err;
 903
 904         get_online_cpus();
 905         mutex_lock(&watchdog_proc_mutex);
 906
 907         if (watchdog_suspended) {
 908                 /* no parameter changes allowed while watchdog is suspended */
 909                 err = -EAGAIN;
 910                 goto out;
 911         }
 912
 913         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
 914         if (!err && write) {
 915                 /* Remove impossible cpus to keep sysctl output cleaner. */
 916                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
 917                             cpu_possible_mask);
 918
 919                 if (watchdog_running) {
 920                         /*
 921                          * Failure would be due to being unable to allocate
 922                          * a temporary cpumask, so we are likely not in a
 923                          * position to do much else to make things better.
 924                          */
 925                         if (watchdog_update_cpus() != 0)
 926                                 pr_err("cpumask update failed\n");
 927                 }
 928
 929                 watchdog_nmi_reconfigure();
 930         }
 931 out:
 932         mutex_unlock(&watchdog_proc_mutex);
 933         put_online_cpus();
 934         return err;
 935 }
 936
 937 #endif /* CONFIG_SYSCTL */
 938
 939 void __init lockup_detector_init(void)
 940 {
 941         set_sample_period();
 942
 943 #ifdef CONFIG_NO_HZ_FULL
 944         if (tick_nohz_full_enabled()) {
 945                 pr_info("Disabling watchdog on nohz_full cores by default\n");
 946                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
 947         } else
 948                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 949 #else
 950         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
 951 #endif
 952
 953         if (watchdog_enabled)
 954                 watchdog_enable_all_cpus();
 955 }