ARM: dtsi: rk3228: add psci support
[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/delay.h>
19 #include <linux/freezer.h>
20 #include <linux/kthread.h>
21 #include <linux/lockdep.h>
22 #include <linux/notifier.h>
23 #include <linux/module.h>
24 #include <linux/sysctl.h>
25 #include <linux/smpboot.h>
26 #include <linux/sched/rt.h>
27
28 #include <asm/irq_regs.h>
29 #include <linux/kvm_para.h>
30 #include <linux/perf_event.h>
31
32 int watchdog_enabled = 1;
33 int __read_mostly watchdog_thresh = 10;
34 static int __read_mostly watchdog_disabled;
35 static u64 __read_mostly sample_period;
36
37 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
38 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
39 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
40 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
41 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
43 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
45 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
46 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
47 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
48 #endif
49 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
50 static cpumask_t __read_mostly watchdog_cpus;
51 #endif
52 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
53 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
54 #endif
55
56 /* boot commands */
57 /*
58  * Should we panic when a soft-lockup or hard-lockup occurs:
59  */
60 #ifdef CONFIG_HARDLOCKUP_DETECTOR
61 static int hardlockup_panic =
62                         CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
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 = 0;
72         return 1;
73 }
74 __setup("nmi_watchdog=", hardlockup_panic_setup);
75 #endif
76
77 unsigned int __read_mostly softlockup_panic =
78                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
79
80 static int __init softlockup_panic_setup(char *str)
81 {
82         softlockup_panic = simple_strtoul(str, NULL, 0);
83
84         return 1;
85 }
86 __setup("softlockup_panic=", softlockup_panic_setup);
87
88 static int __init nowatchdog_setup(char *str)
89 {
90         watchdog_enabled = 0;
91         return 1;
92 }
93 __setup("nowatchdog", nowatchdog_setup);
94
95 /* deprecated */
96 static int __init nosoftlockup_setup(char *str)
97 {
98         watchdog_enabled = 0;
99         return 1;
100 }
101 __setup("nosoftlockup", nosoftlockup_setup);
102 /*  */
103
104 /*
105  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
106  * lockups can have false positives under extreme conditions. So we generally
107  * want a higher threshold for soft lockups than for hard lockups. So we couple
108  * the thresholds with a factor: we make the soft threshold twice the amount of
109  * time the hard threshold is.
110  */
111 static int get_softlockup_thresh(void)
112 {
113         return watchdog_thresh * 2;
114 }
115
116 /*
117  * Returns seconds, approximately.  We don't need nanosecond
118  * resolution, and we don't need to waste time with a big divide when
119  * 2^30ns == 1.074s.
120  */
121 static unsigned long get_timestamp(void)
122 {
123         return local_clock() >> 30LL;  /* 2^30 ~= 10^9 */
124 }
125
126 static void set_sample_period(void)
127 {
128         /*
129          * convert watchdog_thresh from seconds to ns
130          * the divide by 5 is to give hrtimer several chances (two
131          * or three with the current relation between the soft
132          * and hard thresholds) to increment before the
133          * hardlockup detector generates a warning
134          */
135         sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
136 }
137
138 /* Commands for resetting the watchdog */
139 static void __touch_watchdog(void)
140 {
141         __this_cpu_write(watchdog_touch_ts, get_timestamp());
142 }
143
144 void touch_softlockup_watchdog(void)
145 {
146         __this_cpu_write(watchdog_touch_ts, 0);
147 }
148 EXPORT_SYMBOL(touch_softlockup_watchdog);
149
150 void touch_all_softlockup_watchdogs(void)
151 {
152         int cpu;
153
154         /*
155          * this is done lockless
156          * do we care if a 0 races with a timestamp?
157          * all it means is the softlock check starts one cycle later
158          */
159         for_each_online_cpu(cpu)
160                 per_cpu(watchdog_touch_ts, cpu) = 0;
161 }
162
163 #ifdef CONFIG_HARDLOCKUP_DETECTOR
164 void touch_nmi_watchdog(void)
165 {
166         if (watchdog_enabled) {
167                 unsigned cpu;
168
169                 for_each_present_cpu(cpu) {
170                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
171                                 per_cpu(watchdog_nmi_touch, cpu) = true;
172                 }
173         }
174         touch_softlockup_watchdog();
175 }
176 EXPORT_SYMBOL(touch_nmi_watchdog);
177
178 #endif
179
180 void touch_softlockup_watchdog_sync(void)
181 {
182         __raw_get_cpu_var(softlockup_touch_sync) = true;
183         __raw_get_cpu_var(watchdog_touch_ts) = 0;
184 }
185
186 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
187 /* watchdog detector functions */
188 static int is_hardlockup(void)
189 {
190         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
191
192         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
193                 return 1;
194
195         __this_cpu_write(hrtimer_interrupts_saved, hrint);
196         return 0;
197 }
198 #endif
199
200 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
201 static unsigned int watchdog_next_cpu(unsigned int cpu)
202 {
203         cpumask_t cpus = watchdog_cpus;
204         unsigned int next_cpu;
205
206         next_cpu = cpumask_next(cpu, &cpus);
207         if (next_cpu >= nr_cpu_ids)
208                 next_cpu = cpumask_first(&cpus);
209
210         if (next_cpu == cpu)
211                 return nr_cpu_ids;
212
213         return next_cpu;
214 }
215
216 static int is_hardlockup_other_cpu(unsigned int cpu)
217 {
218         unsigned long hrint = per_cpu(hrtimer_interrupts, cpu);
219
220         if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
221                 return 1;
222
223         per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
224         return 0;
225 }
226
227 static void watchdog_check_hardlockup_other_cpu(void)
228 {
229         unsigned int next_cpu;
230
231         /*
232          * Test for hardlockups every 3 samples.  The sample period is
233          *  watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over
234          *  watchdog_thresh (over by 20%).
235          */
236         if (__this_cpu_read(hrtimer_interrupts) % 3 != 0)
237                 return;
238
239         /* check for a hardlockup on the next cpu */
240         next_cpu = watchdog_next_cpu(smp_processor_id());
241         if (next_cpu >= nr_cpu_ids)
242                 return;
243
244         smp_rmb();
245
246         if (per_cpu(watchdog_nmi_touch, next_cpu) == true) {
247                 per_cpu(watchdog_nmi_touch, next_cpu) = false;
248                 return;
249         }
250
251         if (is_hardlockup_other_cpu(next_cpu)) {
252                 /* only warn once */
253                 if (per_cpu(hard_watchdog_warn, next_cpu) == true)
254                         return;
255
256                 if (hardlockup_panic)
257                         panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu);
258                 else
259                         WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu);
260
261                 per_cpu(hard_watchdog_warn, next_cpu) = true;
262         } else {
263                 per_cpu(hard_watchdog_warn, next_cpu) = false;
264         }
265 }
266 #else
267 static inline void watchdog_check_hardlockup_other_cpu(void) { return; }
268 #endif
269
270 static int is_softlockup(unsigned long touch_ts)
271 {
272         unsigned long now = get_timestamp();
273
274         /* Warn about unreasonable delays: */
275         if (time_after(now, touch_ts + get_softlockup_thresh()))
276                 return now - touch_ts;
277
278         return 0;
279 }
280
281 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
282
283 static struct perf_event_attr wd_hw_attr = {
284         .type           = PERF_TYPE_HARDWARE,
285         .config         = PERF_COUNT_HW_CPU_CYCLES,
286         .size           = sizeof(struct perf_event_attr),
287         .pinned         = 1,
288         .disabled       = 1,
289 };
290
291 /* Callback function for perf event subsystem */
292 static void watchdog_overflow_callback(struct perf_event *event,
293                  struct perf_sample_data *data,
294                  struct pt_regs *regs)
295 {
296         /* Ensure the watchdog never gets throttled */
297         event->hw.interrupts = 0;
298
299         if (__this_cpu_read(watchdog_nmi_touch) == true) {
300                 __this_cpu_write(watchdog_nmi_touch, false);
301                 return;
302         }
303
304         /* check for a hardlockup
305          * This is done by making sure our timer interrupt
306          * is incrementing.  The timer interrupt should have
307          * fired multiple times before we overflow'd.  If it hasn't
308          * then this is a good indication the cpu is stuck
309          */
310         if (is_hardlockup()) {
311                 int this_cpu = smp_processor_id();
312
313                 /* only print hardlockups once */
314                 if (__this_cpu_read(hard_watchdog_warn) == true)
315                         return;
316
317                 if (hardlockup_panic)
318                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
319                 else
320                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
321
322                 __this_cpu_write(hard_watchdog_warn, true);
323                 return;
324         }
325
326         __this_cpu_write(hard_watchdog_warn, false);
327         return;
328 }
329 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
330
331 static void watchdog_interrupt_count(void)
332 {
333         __this_cpu_inc(hrtimer_interrupts);
334 }
335
336 static int watchdog_nmi_enable(unsigned int cpu);
337 static void watchdog_nmi_disable(unsigned int cpu);
338
339 /* watchdog kicker functions */
340 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
341 {
342         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
343         struct pt_regs *regs = get_irq_regs();
344         int duration;
345
346         /* kick the hardlockup detector */
347         watchdog_interrupt_count();
348
349         /* test for hardlockups on the next cpu */
350         watchdog_check_hardlockup_other_cpu();
351
352         /* kick the softlockup detector */
353         wake_up_process(__this_cpu_read(softlockup_watchdog));
354
355         /* .. and repeat */
356         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
357
358         if (touch_ts == 0) {
359                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
360                         /*
361                          * If the time stamp was touched atomically
362                          * make sure the scheduler tick is up to date.
363                          */
364                         __this_cpu_write(softlockup_touch_sync, false);
365                         sched_clock_tick();
366                 }
367
368                 /* Clear the guest paused flag on watchdog reset */
369                 kvm_check_and_clear_guest_paused();
370                 __touch_watchdog();
371                 return HRTIMER_RESTART;
372         }
373
374         /* check for a softlockup
375          * This is done by making sure a high priority task is
376          * being scheduled.  The task touches the watchdog to
377          * indicate it is getting cpu time.  If it hasn't then
378          * this is a good indication some task is hogging the cpu
379          */
380         duration = is_softlockup(touch_ts);
381         if (unlikely(duration)) {
382                 /*
383                  * If a virtual machine is stopped by the host it can look to
384                  * the watchdog like a soft lockup, check to see if the host
385                  * stopped the vm before we issue the warning
386                  */
387                 if (kvm_check_and_clear_guest_paused())
388                         return HRTIMER_RESTART;
389
390                 /* only warn once */
391                 if (__this_cpu_read(soft_watchdog_warn) == true)
392                         return HRTIMER_RESTART;
393
394                 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
395                         smp_processor_id(), duration,
396                         current->comm, task_pid_nr(current));
397                 print_modules();
398                 print_irqtrace_events(current);
399                 if (regs)
400                         show_regs(regs);
401                 else
402                         dump_stack();
403
404                 if (softlockup_panic)
405                         panic("softlockup: hung tasks");
406                 __this_cpu_write(soft_watchdog_warn, true);
407         } else
408                 __this_cpu_write(soft_watchdog_warn, false);
409
410         return HRTIMER_RESTART;
411 }
412
413 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
414 {
415         struct sched_param param = { .sched_priority = prio };
416
417         sched_setscheduler(current, policy, &param);
418 }
419
420 static void watchdog_enable(unsigned int cpu)
421 {
422         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
423
424         /* kick off the timer for the hardlockup detector */
425         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
426         hrtimer->function = watchdog_timer_fn;
427
428         if (!watchdog_enabled) {
429                 kthread_park(current);
430                 return;
431         }
432
433         /* Enable the perf event */
434         watchdog_nmi_enable(cpu);
435
436         /* done here because hrtimer_start can only pin to smp_processor_id() */
437         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
438                       HRTIMER_MODE_REL_PINNED);
439
440         /* initialize timestamp */
441         watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
442         __touch_watchdog();
443 }
444
445 static void watchdog_disable(unsigned int cpu)
446 {
447         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
448
449         watchdog_set_prio(SCHED_NORMAL, 0);
450         hrtimer_cancel(hrtimer);
451         /* disable the perf event */
452         watchdog_nmi_disable(cpu);
453 }
454
455 static int watchdog_should_run(unsigned int cpu)
456 {
457         return __this_cpu_read(hrtimer_interrupts) !=
458                 __this_cpu_read(soft_lockup_hrtimer_cnt);
459 }
460
461 /*
462  * The watchdog thread function - touches the timestamp.
463  *
464  * It only runs once every sample_period seconds (4 seconds by
465  * default) to reset the softlockup timestamp. If this gets delayed
466  * for more than 2*watchdog_thresh seconds then the debug-printout
467  * triggers in watchdog_timer_fn().
468  */
469 static void watchdog(unsigned int cpu)
470 {
471         __this_cpu_write(soft_lockup_hrtimer_cnt,
472                          __this_cpu_read(hrtimer_interrupts));
473         __touch_watchdog();
474 }
475
476 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
477 /*
478  * People like the simple clean cpu node info on boot.
479  * Reduce the watchdog noise by only printing messages
480  * that are different from what cpu0 displayed.
481  */
482 static unsigned long cpu0_err;
483
484 static int watchdog_nmi_enable(unsigned int cpu)
485 {
486         struct perf_event_attr *wd_attr;
487         struct perf_event *event = per_cpu(watchdog_ev, cpu);
488
489         /* is it already setup and enabled? */
490         if (event && event->state > PERF_EVENT_STATE_OFF)
491                 goto out;
492
493         /* it is setup but not enabled */
494         if (event != NULL)
495                 goto out_enable;
496
497         wd_attr = &wd_hw_attr;
498         wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
499
500         /* Try to register using hardware perf events */
501         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
502
503         /* save cpu0 error for future comparision */
504         if (cpu == 0 && IS_ERR(event))
505                 cpu0_err = PTR_ERR(event);
506
507         if (!IS_ERR(event)) {
508                 /* only print for cpu0 or different than cpu0 */
509                 if (cpu == 0 || cpu0_err)
510                         pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
511                 goto out_save;
512         }
513
514         /* skip displaying the same error again */
515         if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
516                 return PTR_ERR(event);
517
518         /* vary the KERN level based on the returned errno */
519         if (PTR_ERR(event) == -EOPNOTSUPP)
520                 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
521         else if (PTR_ERR(event) == -ENOENT)
522                 pr_warning("disabled (cpu%i): hardware events not enabled\n",
523                          cpu);
524         else
525                 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
526                         cpu, PTR_ERR(event));
527         return PTR_ERR(event);
528
529         /* success path */
530 out_save:
531         per_cpu(watchdog_ev, cpu) = event;
532 out_enable:
533         perf_event_enable(per_cpu(watchdog_ev, cpu));
534 out:
535         return 0;
536 }
537
538 static void watchdog_nmi_disable(unsigned int cpu)
539 {
540         struct perf_event *event = per_cpu(watchdog_ev, cpu);
541
542         if (event) {
543                 perf_event_disable(event);
544                 per_cpu(watchdog_ev, cpu) = NULL;
545
546                 /* should be in cleanup, but blocks oprofile */
547                 perf_event_release_kernel(event);
548         }
549         return;
550 }
551 #else
552 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
553 static int watchdog_nmi_enable(unsigned int cpu)
554 {
555         /*
556          * The new cpu will be marked online before the first hrtimer interrupt
557          * runs on it.  If another cpu tests for a hardlockup on the new cpu
558          * before it has run its first hrtimer, it will get a false positive.
559          * Touch the watchdog on the new cpu to delay the first check for at
560          * least 3 sampling periods to guarantee one hrtimer has run on the new
561          * cpu.
562          */
563         per_cpu(watchdog_nmi_touch, cpu) = true;
564         smp_wmb();
565         cpumask_set_cpu(cpu, &watchdog_cpus);
566         return 0;
567 }
568
569 static void watchdog_nmi_disable(unsigned int cpu)
570 {
571         unsigned int next_cpu = watchdog_next_cpu(cpu);
572
573         /*
574          * Offlining this cpu will cause the cpu before this one to start
575          * checking the one after this one.  If this cpu just finished checking
576          * the next cpu and updating hrtimer_interrupts_saved, and then the
577          * previous cpu checks it within one sample period, it will trigger a
578          * false positive.  Touch the watchdog on the next cpu to prevent it.
579          */
580         if (next_cpu < nr_cpu_ids)
581                 per_cpu(watchdog_nmi_touch, next_cpu) = true;
582         smp_wmb();
583         cpumask_clear_cpu(cpu, &watchdog_cpus);
584 }
585 #else
586 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
587 static void watchdog_nmi_disable(unsigned int cpu) { return; }
588 #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */
589 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
590
591 /* prepare/enable/disable routines */
592 /* sysctl functions */
593 #ifdef CONFIG_SYSCTL
594 static void watchdog_enable_all_cpus(void)
595 {
596         unsigned int cpu;
597
598         if (watchdog_disabled) {
599                 watchdog_disabled = 0;
600                 for_each_online_cpu(cpu)
601                         kthread_unpark(per_cpu(softlockup_watchdog, cpu));
602         }
603 }
604
605 static void watchdog_disable_all_cpus(void)
606 {
607         unsigned int cpu;
608
609         if (!watchdog_disabled) {
610                 watchdog_disabled = 1;
611                 for_each_online_cpu(cpu)
612                         kthread_park(per_cpu(softlockup_watchdog, cpu));
613         }
614 }
615
616 /*
617  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
618  */
619
620 int proc_dowatchdog(struct ctl_table *table, int write,
621                     void __user *buffer, size_t *lenp, loff_t *ppos)
622 {
623         int ret;
624
625         if (watchdog_disabled < 0)
626                 return -ENODEV;
627
628         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
629         if (ret || !write)
630                 return ret;
631
632         set_sample_period();
633         /*
634          * Watchdog threads shouldn't be enabled if they are
635          * disabled. The 'watchdog_disabled' variable check in
636          * watchdog_*_all_cpus() function takes care of this.
637          */
638         if (watchdog_enabled && watchdog_thresh)
639                 watchdog_enable_all_cpus();
640         else
641                 watchdog_disable_all_cpus();
642
643         return ret;
644 }
645 #endif /* CONFIG_SYSCTL */
646
647 static struct smp_hotplug_thread watchdog_threads = {
648         .store                  = &softlockup_watchdog,
649         .thread_should_run      = watchdog_should_run,
650         .thread_fn              = watchdog,
651         .thread_comm            = "watchdog/%u",
652         .setup                  = watchdog_enable,
653         .park                   = watchdog_disable,
654         .unpark                 = watchdog_enable,
655 };
656
657 void __init lockup_detector_init(void)
658 {
659         set_sample_period();
660         if (smpboot_register_percpu_thread(&watchdog_threads)) {
661                 pr_err("Failed to create watchdog threads, disabled\n");
662                 watchdog_disabled = -ENODEV;
663         }
664 }