watchdog: implement error handling in update_watchdog_all_cpus() and callers
[firefly-linux-kernel-4.4.55.git] / 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
24 #include <asm/irq_regs.h>
25 #include <linux/kvm_para.h>
26 #include <linux/perf_event.h>
27 #include <linux/kthread.h>
28
29 /*
30  * The run state of the lockup detectors is controlled by the content of the
31  * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
32  * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
33  *
34  * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
35  * are variables that are only used as an 'interface' between the parameters
36  * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
37  * 'watchdog_thresh' variable is handled differently because its value is not
38  * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
39  * is equal zero.
40  */
41 #define NMI_WATCHDOG_ENABLED_BIT   0
42 #define SOFT_WATCHDOG_ENABLED_BIT  1
43 #define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
44 #define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
45
46 static DEFINE_MUTEX(watchdog_proc_mutex);
47
48 #ifdef CONFIG_HARDLOCKUP_DETECTOR
49 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
50 #else
51 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
52 #endif
53 int __read_mostly nmi_watchdog_enabled;
54 int __read_mostly soft_watchdog_enabled;
55 int __read_mostly watchdog_user_enabled;
56 int __read_mostly watchdog_thresh = 10;
57
58 #ifdef CONFIG_SMP
59 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
60 #else
61 #define sysctl_softlockup_all_cpu_backtrace 0
62 #endif
63 static struct cpumask watchdog_cpumask __read_mostly;
64 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
65
66 /* Helper for online, unparked cpus. */
67 #define for_each_watchdog_cpu(cpu) \
68         for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
69
70 /*
71  * The 'watchdog_running' variable is set to 1 when the watchdog threads
72  * are registered/started and is set to 0 when the watchdog threads are
73  * unregistered/stopped, so it is an indicator whether the threads exist.
74  */
75 static int __read_mostly watchdog_running;
76 /*
77  * If a subsystem has a need to deactivate the watchdog temporarily, it
78  * can use the suspend/resume interface to achieve this. The content of
79  * the 'watchdog_suspended' variable reflects this state. Existing threads
80  * are parked/unparked by the lockup_detector_{suspend|resume} functions
81  * (see comment blocks pertaining to those functions for further details).
82  *
83  * 'watchdog_suspended' also prevents threads from being registered/started
84  * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
85  * of 'watchdog_running' cannot change while the watchdog is deactivated
86  * temporarily (see related code in 'proc' handlers).
87  */
88 static int __read_mostly watchdog_suspended;
89
90 static u64 __read_mostly sample_period;
91
92 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
93 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
94 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
95 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
96 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
97 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
98 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
99 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
100 #ifdef CONFIG_HARDLOCKUP_DETECTOR
101 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
102 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
103 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
104 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
105 #endif
106 static unsigned long soft_lockup_nmi_warn;
107
108 /* boot commands */
109 /*
110  * Should we panic when a soft-lockup or hard-lockup occurs:
111  */
112 #ifdef CONFIG_HARDLOCKUP_DETECTOR
113 static int hardlockup_panic =
114                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
115 /*
116  * We may not want to enable hard lockup detection by default in all cases,
117  * for example when running the kernel as a guest on a hypervisor. In these
118  * cases this function can be called to disable hard lockup detection. This
119  * function should only be executed once by the boot processor before the
120  * kernel command line parameters are parsed, because otherwise it is not
121  * possible to override this in hardlockup_panic_setup().
122  */
123 void hardlockup_detector_disable(void)
124 {
125         watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
126 }
127
128 static int __init hardlockup_panic_setup(char *str)
129 {
130         if (!strncmp(str, "panic", 5))
131                 hardlockup_panic = 1;
132         else if (!strncmp(str, "nopanic", 7))
133                 hardlockup_panic = 0;
134         else if (!strncmp(str, "0", 1))
135                 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
136         else if (!strncmp(str, "1", 1))
137                 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
138         return 1;
139 }
140 __setup("nmi_watchdog=", hardlockup_panic_setup);
141 #endif
142
143 unsigned int __read_mostly softlockup_panic =
144                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
145
146 static int __init softlockup_panic_setup(char *str)
147 {
148         softlockup_panic = simple_strtoul(str, NULL, 0);
149
150         return 1;
151 }
152 __setup("softlockup_panic=", softlockup_panic_setup);
153
154 static int __init nowatchdog_setup(char *str)
155 {
156         watchdog_enabled = 0;
157         return 1;
158 }
159 __setup("nowatchdog", nowatchdog_setup);
160
161 static int __init nosoftlockup_setup(char *str)
162 {
163         watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
164         return 1;
165 }
166 __setup("nosoftlockup", nosoftlockup_setup);
167
168 #ifdef CONFIG_SMP
169 static int __init softlockup_all_cpu_backtrace_setup(char *str)
170 {
171         sysctl_softlockup_all_cpu_backtrace =
172                 !!simple_strtol(str, NULL, 0);
173         return 1;
174 }
175 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
176 #endif
177
178 /*
179  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
180  * lockups can have false positives under extreme conditions. So we generally
181  * want a higher threshold for soft lockups than for hard lockups. So we couple
182  * the thresholds with a factor: we make the soft threshold twice the amount of
183  * time the hard threshold is.
184  */
185 static int get_softlockup_thresh(void)
186 {
187         return watchdog_thresh * 2;
188 }
189
190 /*
191  * Returns seconds, approximately.  We don't need nanosecond
192  * resolution, and we don't need to waste time with a big divide when
193  * 2^30ns == 1.074s.
194  */
195 static unsigned long get_timestamp(void)
196 {
197         return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
198 }
199
200 static void set_sample_period(void)
201 {
202         /*
203          * convert watchdog_thresh from seconds to ns
204          * the divide by 5 is to give hrtimer several chances (two
205          * or three with the current relation between the soft
206          * and hard thresholds) to increment before the
207          * hardlockup detector generates a warning
208          */
209         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
210 }
211
212 /* Commands for resetting the watchdog */
213 static void __touch_watchdog(void)
214 {
215         __this_cpu_write(watchdog_touch_ts, get_timestamp());
216 }
217
218 void touch_softlockup_watchdog(void)
219 {
220         /*
221          * Preemption can be enabled.  It doesn't matter which CPU's timestamp
222          * gets zeroed here, so use the raw_ operation.
223          */
224         raw_cpu_write(watchdog_touch_ts, 0);
225 }
226 EXPORT_SYMBOL(touch_softlockup_watchdog);
227
228 void touch_all_softlockup_watchdogs(void)
229 {
230         int cpu;
231
232         /*
233          * this is done lockless
234          * do we care if a 0 races with a timestamp?
235          * all it means is the softlock check starts one cycle later
236          */
237         for_each_watchdog_cpu(cpu)
238                 per_cpu(watchdog_touch_ts, cpu) = 0;
239 }
240
241 #ifdef CONFIG_HARDLOCKUP_DETECTOR
242 void touch_nmi_watchdog(void)
243 {
244         /*
245          * Using __raw here because some code paths have
246          * preemption enabled.  If preemption is enabled
247          * then interrupts should be enabled too, in which
248          * case we shouldn't have to worry about the watchdog
249          * going off.
250          */
251         raw_cpu_write(watchdog_nmi_touch, true);
252         touch_softlockup_watchdog();
253 }
254 EXPORT_SYMBOL(touch_nmi_watchdog);
255
256 #endif
257
258 void touch_softlockup_watchdog_sync(void)
259 {
260         __this_cpu_write(softlockup_touch_sync, true);
261         __this_cpu_write(watchdog_touch_ts, 0);
262 }
263
264 #ifdef CONFIG_HARDLOCKUP_DETECTOR
265 /* watchdog detector functions */
266 static bool is_hardlockup(void)
267 {
268         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
269
270         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
271                 return true;
272
273         __this_cpu_write(hrtimer_interrupts_saved, hrint);
274         return false;
275 }
276 #endif
277
278 static int is_softlockup(unsigned long touch_ts)
279 {
280         unsigned long now = get_timestamp();
281
282         if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) {
283                 /* Warn about unreasonable delays. */
284                 if (time_after(now, touch_ts + get_softlockup_thresh()))
285                         return now - touch_ts;
286         }
287         return 0;
288 }
289
290 #ifdef CONFIG_HARDLOCKUP_DETECTOR
291
292 static struct perf_event_attr wd_hw_attr = {
293         .type           = PERF_TYPE_HARDWARE,
294         .config         = PERF_COUNT_HW_CPU_CYCLES,
295         .size           = sizeof(struct perf_event_attr),
296         .pinned         = 1,
297         .disabled       = 1,
298 };
299
300 /* Callback function for perf event subsystem */
301 static void watchdog_overflow_callback(struct perf_event *event,
302                  struct perf_sample_data *data,
303                  struct pt_regs *regs)
304 {
305         /* Ensure the watchdog never gets throttled */
306         event->hw.interrupts = 0;
307
308         if (__this_cpu_read(watchdog_nmi_touch) == true) {
309                 __this_cpu_write(watchdog_nmi_touch, false);
310                 return;
311         }
312
313         /* check for a hardlockup
314          * This is done by making sure our timer interrupt
315          * is incrementing.  The timer interrupt should have
316          * fired multiple times before we overflow'd.  If it hasn't
317          * then this is a good indication the cpu is stuck
318          */
319         if (is_hardlockup()) {
320                 int this_cpu = smp_processor_id();
321
322                 /* only print hardlockups once */
323                 if (__this_cpu_read(hard_watchdog_warn) == true)
324                         return;
325
326                 if (hardlockup_panic)
327                         panic("Watchdog detected hard LOCKUP on cpu %d",
328                               this_cpu);
329                 else
330                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
331                              this_cpu);
332
333                 __this_cpu_write(hard_watchdog_warn, true);
334                 return;
335         }
336
337         __this_cpu_write(hard_watchdog_warn, false);
338         return;
339 }
340 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
341
342 static void watchdog_interrupt_count(void)
343 {
344         __this_cpu_inc(hrtimer_interrupts);
345 }
346
347 static int watchdog_nmi_enable(unsigned int cpu);
348 static void watchdog_nmi_disable(unsigned int cpu);
349
350 static int watchdog_enable_all_cpus(void);
351 static void watchdog_disable_all_cpus(void);
352
353 /* watchdog kicker functions */
354 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
355 {
356         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
357         struct pt_regs *regs = get_irq_regs();
358         int duration;
359         int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
360
361         /* kick the hardlockup detector */
362         watchdog_interrupt_count();
363
364         /* kick the softlockup detector */
365         wake_up_process(__this_cpu_read(softlockup_watchdog));
366
367         /* .. and repeat */
368         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
369
370         if (touch_ts == 0) {
371                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
372                         /*
373                          * If the time stamp was touched atomically
374                          * make sure the scheduler tick is up to date.
375                          */
376                         __this_cpu_write(softlockup_touch_sync, false);
377                         sched_clock_tick();
378                 }
379
380                 /* Clear the guest paused flag on watchdog reset */
381                 kvm_check_and_clear_guest_paused();
382                 __touch_watchdog();
383                 return HRTIMER_RESTART;
384         }
385
386         /* check for a softlockup
387          * This is done by making sure a high priority task is
388          * being scheduled.  The task touches the watchdog to
389          * indicate it is getting cpu time.  If it hasn't then
390          * this is a good indication some task is hogging the cpu
391          */
392         duration = is_softlockup(touch_ts);
393         if (unlikely(duration)) {
394                 /*
395                  * If a virtual machine is stopped by the host it can look to
396                  * the watchdog like a soft lockup, check to see if the host
397                  * stopped the vm before we issue the warning
398                  */
399                 if (kvm_check_and_clear_guest_paused())
400                         return HRTIMER_RESTART;
401
402                 /* only warn once */
403                 if (__this_cpu_read(soft_watchdog_warn) == true) {
404                         /*
405                          * When multiple processes are causing softlockups the
406                          * softlockup detector only warns on the first one
407                          * because the code relies on a full quiet cycle to
408                          * re-arm.  The second process prevents the quiet cycle
409                          * and never gets reported.  Use task pointers to detect
410                          * this.
411                          */
412                         if (__this_cpu_read(softlockup_task_ptr_saved) !=
413                             current) {
414                                 __this_cpu_write(soft_watchdog_warn, false);
415                                 __touch_watchdog();
416                         }
417                         return HRTIMER_RESTART;
418                 }
419
420                 if (softlockup_all_cpu_backtrace) {
421                         /* Prevent multiple soft-lockup reports if one cpu is already
422                          * engaged in dumping cpu back traces
423                          */
424                         if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
425                                 /* Someone else will report us. Let's give up */
426                                 __this_cpu_write(soft_watchdog_warn, true);
427                                 return HRTIMER_RESTART;
428                         }
429                 }
430
431                 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
432                         smp_processor_id(), duration,
433                         current->comm, task_pid_nr(current));
434                 __this_cpu_write(softlockup_task_ptr_saved, current);
435                 print_modules();
436                 print_irqtrace_events(current);
437                 if (regs)
438                         show_regs(regs);
439                 else
440                         dump_stack();
441
442                 if (softlockup_all_cpu_backtrace) {
443                         /* Avoid generating two back traces for current
444                          * given that one is already made above
445                          */
446                         trigger_allbutself_cpu_backtrace();
447
448                         clear_bit(0, &soft_lockup_nmi_warn);
449                         /* Barrier to sync with other cpus */
450                         smp_mb__after_atomic();
451                 }
452
453                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
454                 if (softlockup_panic)
455                         panic("softlockup: hung tasks");
456                 __this_cpu_write(soft_watchdog_warn, true);
457         } else
458                 __this_cpu_write(soft_watchdog_warn, false);
459
460         return HRTIMER_RESTART;
461 }
462
463 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
464 {
465         struct sched_param param = { .sched_priority = prio };
466
467         sched_setscheduler(current, policy, &param);
468 }
469
470 static void watchdog_enable(unsigned int cpu)
471 {
472         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
473
474         /* kick off the timer for the hardlockup detector */
475         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
476         hrtimer->function = watchdog_timer_fn;
477
478         /* Enable the perf event */
479         watchdog_nmi_enable(cpu);
480
481         /* done here because hrtimer_start can only pin to smp_processor_id() */
482         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
483                       HRTIMER_MODE_REL_PINNED);
484
485         /* initialize timestamp */
486         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
487         __touch_watchdog();
488 }
489
490 static void watchdog_disable(unsigned int cpu)
491 {
492         struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
493
494         watchdog_set_prio(SCHED_NORMAL, 0);
495         hrtimer_cancel(hrtimer);
496         /* disable the perf event */
497         watchdog_nmi_disable(cpu);
498 }
499
500 static void watchdog_cleanup(unsigned int cpu, bool online)
501 {
502         watchdog_disable(cpu);
503 }
504
505 static int watchdog_should_run(unsigned int cpu)
506 {
507         return __this_cpu_read(hrtimer_interrupts) !=
508                 __this_cpu_read(soft_lockup_hrtimer_cnt);
509 }
510
511 /*
512  * The watchdog thread function - touches the timestamp.
513  *
514  * It only runs once every sample_period seconds (4 seconds by
515  * default) to reset the softlockup timestamp. If this gets delayed
516  * for more than 2*watchdog_thresh seconds then the debug-printout
517  * triggers in watchdog_timer_fn().
518  */
519 static void watchdog(unsigned int cpu)
520 {
521         __this_cpu_write(soft_lockup_hrtimer_cnt,
522                          __this_cpu_read(hrtimer_interrupts));
523         __touch_watchdog();
524
525         /*
526          * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
527          * failure path. Check for failures that can occur asynchronously -
528          * for example, when CPUs are on-lined - and shut down the hardware
529          * perf event on each CPU accordingly.
530          *
531          * The only non-obvious place this bit can be cleared is through
532          * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
533          * pr_info here would be too noisy as it would result in a message
534          * every few seconds if the hardlockup was disabled but the softlockup
535          * enabled.
536          */
537         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
538                 watchdog_nmi_disable(cpu);
539 }
540
541 #ifdef CONFIG_HARDLOCKUP_DETECTOR
542 /*
543  * People like the simple clean cpu node info on boot.
544  * Reduce the watchdog noise by only printing messages
545  * that are different from what cpu0 displayed.
546  */
547 static unsigned long cpu0_err;
548
549 static int watchdog_nmi_enable(unsigned int cpu)
550 {
551         struct perf_event_attr *wd_attr;
552         struct perf_event *event = per_cpu(watchdog_ev, cpu);
553
554         /* nothing to do if the hard lockup detector is disabled */
555         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
556                 goto out;
557
558         /* is it already setup and enabled? */
559         if (event && event->state > PERF_EVENT_STATE_OFF)
560                 goto out;
561
562         /* it is setup but not enabled */
563         if (event != NULL)
564                 goto out_enable;
565
566         wd_attr = &wd_hw_attr;
567         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
568
569         /* Try to register using hardware perf events */
570         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
571
572         /* save cpu0 error for future comparision */
573         if (cpu == 0 && IS_ERR(event))
574                 cpu0_err = PTR_ERR(event);
575
576         if (!IS_ERR(event)) {
577                 /* only print for cpu0 or different than cpu0 */
578                 if (cpu == 0 || cpu0_err)
579                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
580                 goto out_save;
581         }
582
583         /*
584          * Disable the hard lockup detector if _any_ CPU fails to set up
585          * set up the hardware perf event. The watchdog() function checks
586          * the NMI_WATCHDOG_ENABLED bit periodically.
587          *
588          * The barriers are for syncing up watchdog_enabled across all the
589          * cpus, as clear_bit() does not use barriers.
590          */
591         smp_mb__before_atomic();
592         clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
593         smp_mb__after_atomic();
594
595         /* skip displaying the same error again */
596         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
597                 return PTR_ERR(event);
598
599         /* vary the KERN level based on the returned errno */
600         if (PTR_ERR(event) == -EOPNOTSUPP)
601                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
602         else if (PTR_ERR(event) == -ENOENT)
603                 pr_warn("disabled (cpu%i): hardware events not enabled\n",
604                          cpu);
605         else
606                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
607                         cpu, PTR_ERR(event));
608
609         pr_info("Shutting down hard lockup detector on all cpus\n");
610
611         return PTR_ERR(event);
612
613         /* success path */
614 out_save:
615         per_cpu(watchdog_ev, cpu) = event;
616 out_enable:
617         perf_event_enable(per_cpu(watchdog_ev, cpu));
618 out:
619         return 0;
620 }
621
622 static void watchdog_nmi_disable(unsigned int cpu)
623 {
624         struct perf_event *event = per_cpu(watchdog_ev, cpu);
625
626         if (event) {
627                 perf_event_disable(event);
628                 per_cpu(watchdog_ev, cpu) = NULL;
629
630                 /* should be in cleanup, but blocks oprofile */
631                 perf_event_release_kernel(event);
632         }
633         if (cpu == 0) {
634                 /* watchdog_nmi_enable() expects this to be zero initially. */
635                 cpu0_err = 0;
636         }
637 }
638
639 #else
640 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
641 static void watchdog_nmi_disable(unsigned int cpu) { return; }
642 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
643
644 static struct smp_hotplug_thread watchdog_threads = {
645         .store                  = &softlockup_watchdog,
646         .thread_should_run      = watchdog_should_run,
647         .thread_fn              = watchdog,
648         .thread_comm            = "watchdog/%u",
649         .setup                  = watchdog_enable,
650         .cleanup                = watchdog_cleanup,
651         .park                   = watchdog_disable,
652         .unpark                 = watchdog_enable,
653 };
654
655 /*
656  * park all watchdog threads that are specified in 'watchdog_cpumask'
657  */
658 static int watchdog_park_threads(void)
659 {
660         int cpu, ret = 0;
661
662         get_online_cpus();
663         for_each_watchdog_cpu(cpu) {
664                 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
665                 if (ret)
666                         break;
667         }
668         if (ret) {
669                 for_each_watchdog_cpu(cpu)
670                         kthread_unpark(per_cpu(softlockup_watchdog, cpu));
671         }
672         put_online_cpus();
673
674         return ret;
675 }
676
677 /*
678  * unpark all watchdog threads that are specified in 'watchdog_cpumask'
679  */
680 static void watchdog_unpark_threads(void)
681 {
682         int cpu;
683
684         get_online_cpus();
685         for_each_watchdog_cpu(cpu)
686                 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
687         put_online_cpus();
688 }
689
690 /*
691  * Suspend the hard and soft lockup detector by parking the watchdog threads.
692  */
693 int lockup_detector_suspend(void)
694 {
695         int ret = 0;
696
697         mutex_lock(&watchdog_proc_mutex);
698         /*
699          * Multiple suspend requests can be active in parallel (counted by
700          * the 'watchdog_suspended' variable). If the watchdog threads are
701          * running, the first caller takes care that they will be parked.
702          * The state of 'watchdog_running' cannot change while a suspend
703          * request is active (see related code in 'proc' handlers).
704          */
705         if (watchdog_running && !watchdog_suspended)
706                 ret = watchdog_park_threads();
707
708         if (ret == 0)
709                 watchdog_suspended++;
710
711         mutex_unlock(&watchdog_proc_mutex);
712
713         return ret;
714 }
715
716 /*
717  * Resume the hard and soft lockup detector by unparking the watchdog threads.
718  */
719 void lockup_detector_resume(void)
720 {
721         mutex_lock(&watchdog_proc_mutex);
722
723         watchdog_suspended--;
724         /*
725          * The watchdog threads are unparked if they were previously running
726          * and if there is no more active suspend request.
727          */
728         if (watchdog_running && !watchdog_suspended)
729                 watchdog_unpark_threads();
730
731         mutex_unlock(&watchdog_proc_mutex);
732 }
733
734 static int update_watchdog_all_cpus(void)
735 {
736         int ret;
737
738         ret = watchdog_park_threads();
739         if (ret)
740                 return ret;
741
742         watchdog_unpark_threads();
743
744         return 0;
745 }
746
747 static int watchdog_enable_all_cpus(void)
748 {
749         int err = 0;
750
751         if (!watchdog_running) {
752                 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
753                                                              &watchdog_cpumask);
754                 if (err)
755                         pr_err("Failed to create watchdog threads, disabled\n");
756                 else
757                         watchdog_running = 1;
758         } else {
759                 /*
760                  * Enable/disable the lockup detectors or
761                  * change the sample period 'on the fly'.
762                  */
763                 err = update_watchdog_all_cpus();
764
765                 if (err) {
766                         watchdog_disable_all_cpus();
767                         pr_err("Failed to update lockup detectors, disabled\n");
768                 }
769         }
770
771         if (err)
772                 watchdog_enabled = 0;
773
774         return err;
775 }
776
777 static void watchdog_disable_all_cpus(void)
778 {
779         if (watchdog_running) {
780                 watchdog_running = 0;
781                 smpboot_unregister_percpu_thread(&watchdog_threads);
782         }
783 }
784
785 #ifdef CONFIG_SYSCTL
786
787 /*
788  * Update the run state of the lockup detectors.
789  */
790 static int proc_watchdog_update(void)
791 {
792         int err = 0;
793
794         /*
795          * Watchdog threads won't be started if they are already active.
796          * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
797          * care of this. If those threads are already active, the sample
798          * period will be updated and the lockup detectors will be enabled
799          * or disabled 'on the fly'.
800          */
801         if (watchdog_enabled && watchdog_thresh)
802                 err = watchdog_enable_all_cpus();
803         else
804                 watchdog_disable_all_cpus();
805
806         return err;
807
808 }
809
810 /*
811  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
812  *
813  * caller             | table->data points to | 'which' contains the flag(s)
814  * -------------------|-----------------------|-----------------------------
815  * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
816  *                    |                       | with SOFT_WATCHDOG_ENABLED
817  * -------------------|-----------------------|-----------------------------
818  * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
819  * -------------------|-----------------------|-----------------------------
820  * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
821  */
822 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
823                                 void __user *buffer, size_t *lenp, loff_t *ppos)
824 {
825         int err, old, new;
826         int *watchdog_param = (int *)table->data;
827
828         mutex_lock(&watchdog_proc_mutex);
829
830         if (watchdog_suspended) {
831                 /* no parameter changes allowed while watchdog is suspended */
832                 err = -EAGAIN;
833                 goto out;
834         }
835
836         /*
837          * If the parameter is being read return the state of the corresponding
838          * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
839          * run state of the lockup detectors.
840          */
841         if (!write) {
842                 *watchdog_param = (watchdog_enabled & which) != 0;
843                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
844         } else {
845                 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
846                 if (err)
847                         goto out;
848
849                 /*
850                  * There is a race window between fetching the current value
851                  * from 'watchdog_enabled' and storing the new value. During
852                  * this race window, watchdog_nmi_enable() can sneak in and
853                  * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
854                  * The 'cmpxchg' detects this race and the loop retries.
855                  */
856                 do {
857                         old = watchdog_enabled;
858                         /*
859                          * If the parameter value is not zero set the
860                          * corresponding bit(s), else clear it(them).
861                          */
862                         if (*watchdog_param)
863                                 new = old | which;
864                         else
865                                 new = old & ~which;
866                 } while (cmpxchg(&watchdog_enabled, old, new) != old);
867
868                 /*
869                  * Update the run state of the lockup detectors. There is _no_
870                  * need to check the value returned by proc_watchdog_update()
871                  * and to restore the previous value of 'watchdog_enabled' as
872                  * both lockup detectors are disabled if proc_watchdog_update()
873                  * returns an error.
874                  */
875                 err = proc_watchdog_update();
876         }
877 out:
878         mutex_unlock(&watchdog_proc_mutex);
879         return err;
880 }
881
882 /*
883  * /proc/sys/kernel/watchdog
884  */
885 int proc_watchdog(struct ctl_table *table, int write,
886                   void __user *buffer, size_t *lenp, loff_t *ppos)
887 {
888         return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
889                                     table, write, buffer, lenp, ppos);
890 }
891
892 /*
893  * /proc/sys/kernel/nmi_watchdog
894  */
895 int proc_nmi_watchdog(struct ctl_table *table, int write,
896                       void __user *buffer, size_t *lenp, loff_t *ppos)
897 {
898         return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
899                                     table, write, buffer, lenp, ppos);
900 }
901
902 /*
903  * /proc/sys/kernel/soft_watchdog
904  */
905 int proc_soft_watchdog(struct ctl_table *table, int write,
906                         void __user *buffer, size_t *lenp, loff_t *ppos)
907 {
908         return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
909                                     table, write, buffer, lenp, ppos);
910 }
911
912 /*
913  * /proc/sys/kernel/watchdog_thresh
914  */
915 int proc_watchdog_thresh(struct ctl_table *table, int write,
916                          void __user *buffer, size_t *lenp, loff_t *ppos)
917 {
918         int err, old;
919
920         mutex_lock(&watchdog_proc_mutex);
921
922         if (watchdog_suspended) {
923                 /* no parameter changes allowed while watchdog is suspended */
924                 err = -EAGAIN;
925                 goto out;
926         }
927
928         old = ACCESS_ONCE(watchdog_thresh);
929         err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
930
931         if (err || !write)
932                 goto out;
933
934         /*
935          * Update the sample period. Restore on failure.
936          */
937         set_sample_period();
938         err = proc_watchdog_update();
939         if (err) {
940                 watchdog_thresh = old;
941                 set_sample_period();
942         }
943 out:
944         mutex_unlock(&watchdog_proc_mutex);
945         return err;
946 }
947
948 /*
949  * The cpumask is the mask of possible cpus that the watchdog can run
950  * on, not the mask of cpus it is actually running on.  This allows the
951  * user to specify a mask that will include cpus that have not yet
952  * been brought online, if desired.
953  */
954 int proc_watchdog_cpumask(struct ctl_table *table, int write,
955                           void __user *buffer, size_t *lenp, loff_t *ppos)
956 {
957         int err;
958
959         mutex_lock(&watchdog_proc_mutex);
960
961         if (watchdog_suspended) {
962                 /* no parameter changes allowed while watchdog is suspended */
963                 err = -EAGAIN;
964                 goto out;
965         }
966
967         err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
968         if (!err && write) {
969                 /* Remove impossible cpus to keep sysctl output cleaner. */
970                 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
971                             cpu_possible_mask);
972
973                 if (watchdog_running) {
974                         /*
975                          * Failure would be due to being unable to allocate
976                          * a temporary cpumask, so we are likely not in a
977                          * position to do much else to make things better.
978                          */
979                         if (smpboot_update_cpumask_percpu_thread(
980                                     &watchdog_threads, &watchdog_cpumask) != 0)
981                                 pr_err("cpumask update failed\n");
982                 }
983         }
984 out:
985         mutex_unlock(&watchdog_proc_mutex);
986         return err;
987 }
988
989 #endif /* CONFIG_SYSCTL */
990
991 void __init lockup_detector_init(void)
992 {
993         set_sample_period();
994
995 #ifdef CONFIG_NO_HZ_FULL
996         if (tick_nohz_full_enabled()) {
997                 pr_info("Disabling watchdog on nohz_full cores by default\n");
998                 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
999         } else
1000                 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1001 #else
1002         cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1003 #endif
1004
1005         if (watchdog_enabled)
1006                 watchdog_enable_all_cpus();
1007 }