blob: 5693afd2b8eaea482b2ab7715d10b3e3290f5dbe [file] [log] [blame]
Don Zickus58687ac2010-05-07 17:11:44 -04001/*
2 * Detect hard and soft lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -05006 * 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
Don Zickus58687ac2010-05-07 17:11:44 -04009 * to those contributors as well.
10 */
11
Kefeng Wang5f92a7b2017-07-14 14:49:46 -070012#define pr_fmt(fmt) "watchdog: " fmt
Andrew Morton45019802012-03-23 15:01:55 -070013
Don Zickus58687ac2010-05-07 17:11:44 -040014#include <linux/mm.h>
15#include <linux/cpu.h>
16#include <linux/nmi.h>
17#include <linux/init.h>
Don Zickus58687ac2010-05-07 17:11:44 -040018#include <linux/module.h>
19#include <linux/sysctl.h>
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +000020#include <linux/smpboot.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060021#include <linux/sched/rt.h>
Ingo Molnarae7e81c2017-02-01 18:07:51 +010022#include <uapi/linux/sched/types.h>
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070023#include <linux/tick.h>
Tejun Heo82607adc2015-12-08 11:28:04 -050024#include <linux/workqueue.h>
Ingo Molnare6017572017-02-01 16:36:40 +010025#include <linux/sched/clock.h>
Ingo Molnarb17b0152017-02-08 18:51:35 +010026#include <linux/sched/debug.h>
Don Zickus58687ac2010-05-07 17:11:44 -040027
28#include <asm/irq_regs.h>
Eric B Munson5d1c0f42012-03-10 14:37:28 -050029#include <linux/kvm_para.h>
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -070030#include <linux/kthread.h>
Don Zickus58687ac2010-05-07 17:11:44 -040031
Thomas Gleixner946d1972017-09-12 21:37:01 +020032static DEFINE_MUTEX(watchdog_mutex);
Peter Zijlstraab992dc2015-05-18 11:31:50 +020033
Nicholas Piggin05a4a952017-07-12 14:35:46 -070034#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
35unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED |
36 NMI_WATCHDOG_ENABLED;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070037#else
Babu Moger249e52e2016-12-14 15:06:21 -080038unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070039#endif
Nicholas Piggin05a4a952017-07-12 14:35:46 -070040
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020041int __read_mostly nmi_watchdog_user_enabled;
42int __read_mostly soft_watchdog_user_enabled;
43int __read_mostly watchdog_user_enabled;
44int __read_mostly watchdog_thresh = 10;
45
46struct cpumask watchdog_allowed_mask __read_mostly;
47static bool softlockup_threads_initialized __read_mostly;
48
49struct cpumask watchdog_cpumask __read_mostly;
50unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
51
Nicholas Piggin05a4a952017-07-12 14:35:46 -070052#ifdef CONFIG_HARDLOCKUP_DETECTOR
Nicholas Piggin05a4a952017-07-12 14:35:46 -070053/*
54 * Should we panic when a soft-lockup or hard-lockup occurs:
55 */
56unsigned int __read_mostly hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
58/*
59 * We may not want to enable hard lockup detection by default in all cases,
60 * for example when running the kernel as a guest on a hypervisor. In these
61 * cases this function can be called to disable hard lockup detection. This
62 * function should only be executed once by the boot processor before the
63 * kernel command line parameters are parsed, because otherwise it is not
64 * possible to override this in hardlockup_panic_setup().
65 */
Thomas Gleixner7a355822017-09-12 21:37:02 +020066void __init hardlockup_detector_disable(void)
Nicholas Piggin05a4a952017-07-12 14:35:46 -070067{
68 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
69}
70
71static int __init hardlockup_panic_setup(char *str)
72{
73 if (!strncmp(str, "panic", 5))
74 hardlockup_panic = 1;
75 else if (!strncmp(str, "nopanic", 7))
76 hardlockup_panic = 0;
77 else if (!strncmp(str, "0", 1))
78 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
79 else if (!strncmp(str, "1", 1))
80 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
81 return 1;
82}
83__setup("nmi_watchdog=", hardlockup_panic_setup);
84
Thomas Gleixner368a7e22017-09-12 21:37:07 +020085# ifdef CONFIG_SMP
86int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070087
Thomas Gleixner368a7e22017-09-12 21:37:07 +020088static int __init hardlockup_all_cpu_backtrace_setup(char *str)
89{
90 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
91 return 1;
92}
93__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
94# endif /* CONFIG_SMP */
95#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -070096
Ulrich Obergfellec6a9062015-09-04 15:45:28 -070097/*
Nicholas Piggin05a4a952017-07-12 14:35:46 -070098 * These functions can be overridden if an architecture implements its
99 * own hardlockup detector.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700100 *
101 * watchdog_nmi_enable/disable can be implemented to start and stop when
102 * softlockup watchdog threads start and stop. The arch must select the
103 * SOFTLOCKUP_DETECTOR Kconfig.
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700104 */
105int __weak watchdog_nmi_enable(unsigned int cpu)
106{
107 return 0;
108}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200109
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700110void __weak watchdog_nmi_disable(unsigned int cpu)
111{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200112 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700113}
114
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200115/**
116 * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs
117 * @run: If false stop the watchdogs on all enabled CPUs
118 * If true start the watchdogs on all enabled CPUs
119 *
120 * The core call order is:
121 * watchdog_nmi_reconfigure(false);
122 * update_variables();
123 * watchdog_nmi_reconfigure(true);
124 *
125 * The second call which starts the watchdogs again guarantees that the
126 * following variables are stable across the call.
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200127 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700128 * - watchdog_thresh
129 * - watchdog_cpumask
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200130 *
131 * After the call the variables can be changed again.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700132 */
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200133void __weak watchdog_nmi_reconfigure(bool run) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700134
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700135#ifdef CONFIG_SOFTLOCKUP_DETECTOR
136
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200137/* Global variables, exported for sysctl */
138unsigned int __read_mostly softlockup_panic =
139 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200140
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800141static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400142
143static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
144static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
145static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
146static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400147static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000148static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
149static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700150static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400151static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700152static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400153
Don Zickus58687ac2010-05-07 17:11:44 -0400154static int __init softlockup_panic_setup(char *str)
155{
156 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400157 return 1;
158}
159__setup("softlockup_panic=", softlockup_panic_setup);
160
161static int __init nowatchdog_setup(char *str)
162{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700163 watchdog_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400164 return 1;
165}
166__setup("nowatchdog", nowatchdog_setup);
167
Don Zickus58687ac2010-05-07 17:11:44 -0400168static int __init nosoftlockup_setup(char *str)
169{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700170 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
Don Zickus58687ac2010-05-07 17:11:44 -0400171 return 1;
172}
173__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700174
Aaron Tomlined235872014-06-23 13:22:05 -0700175#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200176int __read_mostly sysctl_softlockup_all_cpu_backtrace;
177
Aaron Tomlined235872014-06-23 13:22:05 -0700178static int __init softlockup_all_cpu_backtrace_setup(char *str)
179{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200180 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700181 return 1;
182}
183__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700184#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400185
Thomas Gleixner941154b2017-09-12 21:37:04 +0200186static void __lockup_detector_cleanup(void);
187
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700188/*
189 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
190 * lockups can have false positives under extreme conditions. So we generally
191 * want a higher threshold for soft lockups than for hard lockups. So we couple
192 * the thresholds with a factor: we make the soft threshold twice the amount of
193 * time the hard threshold is.
194 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200195static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700196{
197 return watchdog_thresh * 2;
198}
Don Zickus58687ac2010-05-07 17:11:44 -0400199
200/*
201 * Returns seconds, approximately. We don't need nanosecond
202 * resolution, and we don't need to waste time with a big divide when
203 * 2^30ns == 1.074s.
204 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900205static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400206{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800207 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400208}
209
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800210static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400211{
212 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700213 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500214 * the divide by 5 is to give hrtimer several chances (two
215 * or three with the current relation between the soft
216 * and hard thresholds) to increment before the
217 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400218 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800219 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200220 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400221}
222
223/* Commands for resetting the watchdog */
224static void __touch_watchdog(void)
225{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900226 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400227}
228
Tejun Heo03e0d462015-12-08 11:28:04 -0500229/**
230 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
231 *
232 * Call when the scheduler may have stalled for legitimate reasons
233 * preventing the watchdog task from executing - e.g. the scheduler
234 * entering idle state. This should only be used for scheduler events.
235 * Use touch_softlockup_watchdog() for everything else.
236 */
237void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400238{
Andrew Morton78611442014-04-18 15:07:12 -0700239 /*
240 * Preemption can be enabled. It doesn't matter which CPU's timestamp
241 * gets zeroed here, so use the raw_ operation.
242 */
243 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400244}
Tejun Heo03e0d462015-12-08 11:28:04 -0500245
246void touch_softlockup_watchdog(void)
247{
248 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500249 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500250}
Ingo Molnar0167c782010-05-13 08:53:33 +0200251EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400252
Don Zickus332fbdb2010-05-07 17:11:45 -0400253void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400254{
255 int cpu;
256
257 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200258 * watchdog_mutex cannpt be taken here, as this might be called
259 * from (soft)interrupt context, so the access to
260 * watchdog_allowed_cpumask might race with a concurrent update.
261 *
262 * The watchdog time stamp can race against a concurrent real
263 * update as well, the only side effect might be a cycle delay for
264 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400265 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200266 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400267 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500268 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400269}
270
Don Zickus58687ac2010-05-07 17:11:44 -0400271void touch_softlockup_watchdog_sync(void)
272{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500273 __this_cpu_write(softlockup_touch_sync, true);
274 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400275}
276
Don Zickus26e09c62010-05-17 18:06:04 -0400277static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400278{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900279 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400280
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800281 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700282 /* Warn about unreasonable delays. */
283 if (time_after(now, touch_ts + get_softlockup_thresh()))
284 return now - touch_ts;
285 }
Don Zickus58687ac2010-05-07 17:11:44 -0400286 return 0;
287}
288
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700289/* watchdog detector functions */
290bool is_hardlockup(void)
291{
292 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
293
294 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
295 return true;
296
297 __this_cpu_write(hrtimer_interrupts_saved, hrint);
298 return false;
299}
300
Don Zickus58687ac2010-05-07 17:11:44 -0400301static void watchdog_interrupt_count(void)
302{
Christoph Lameter909ea962010-12-08 16:22:55 +0100303 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400304}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000305
Don Zickus58687ac2010-05-07 17:11:44 -0400306/* watchdog kicker functions */
307static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
308{
Christoph Lameter909ea962010-12-08 16:22:55 +0100309 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400310 struct pt_regs *regs = get_irq_regs();
311 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700312 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400313
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200314 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800315 return HRTIMER_NORESTART;
316
Don Zickus58687ac2010-05-07 17:11:44 -0400317 /* kick the hardlockup detector */
318 watchdog_interrupt_count();
319
320 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100321 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400322
323 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800324 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400325
326 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100327 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400328 /*
329 * If the time stamp was touched atomically
330 * make sure the scheduler tick is up to date.
331 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100332 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400333 sched_clock_tick();
334 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500335
336 /* Clear the guest paused flag on watchdog reset */
337 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400338 __touch_watchdog();
339 return HRTIMER_RESTART;
340 }
341
342 /* check for a softlockup
343 * This is done by making sure a high priority task is
344 * being scheduled. The task touches the watchdog to
345 * indicate it is getting cpu time. If it hasn't then
346 * this is a good indication some task is hogging the cpu
347 */
Don Zickus26e09c62010-05-17 18:06:04 -0400348 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400349 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500350 /*
351 * If a virtual machine is stopped by the host it can look to
352 * the watchdog like a soft lockup, check to see if the host
353 * stopped the vm before we issue the warning
354 */
355 if (kvm_check_and_clear_guest_paused())
356 return HRTIMER_RESTART;
357
Don Zickus58687ac2010-05-07 17:11:44 -0400358 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700359 if (__this_cpu_read(soft_watchdog_warn) == true) {
360 /*
361 * When multiple processes are causing softlockups the
362 * softlockup detector only warns on the first one
363 * because the code relies on a full quiet cycle to
364 * re-arm. The second process prevents the quiet cycle
365 * and never gets reported. Use task pointers to detect
366 * this.
367 */
368 if (__this_cpu_read(softlockup_task_ptr_saved) !=
369 current) {
370 __this_cpu_write(soft_watchdog_warn, false);
371 __touch_watchdog();
372 }
Don Zickus58687ac2010-05-07 17:11:44 -0400373 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700374 }
Don Zickus58687ac2010-05-07 17:11:44 -0400375
Aaron Tomlined235872014-06-23 13:22:05 -0700376 if (softlockup_all_cpu_backtrace) {
377 /* Prevent multiple soft-lockup reports if one cpu is already
378 * engaged in dumping cpu back traces
379 */
380 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
381 /* Someone else will report us. Let's give up */
382 __this_cpu_write(soft_watchdog_warn, true);
383 return HRTIMER_RESTART;
384 }
385 }
386
Fabian Frederick656c3b72014-08-06 16:04:03 -0700387 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400388 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400389 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700390 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400391 print_modules();
392 print_irqtrace_events(current);
393 if (regs)
394 show_regs(regs);
395 else
396 dump_stack();
397
Aaron Tomlined235872014-06-23 13:22:05 -0700398 if (softlockup_all_cpu_backtrace) {
399 /* Avoid generating two back traces for current
400 * given that one is already made above
401 */
402 trigger_allbutself_cpu_backtrace();
403
404 clear_bit(0, &soft_lockup_nmi_warn);
405 /* Barrier to sync with other cpus */
406 smp_mb__after_atomic();
407 }
408
Josh Hunt69361ee2014-08-08 14:22:31 -0700409 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400410 if (softlockup_panic)
411 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100412 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400413 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100414 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400415
416 return HRTIMER_RESTART;
417}
418
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000419static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400420{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000421 struct sched_param param = { .sched_priority = prio };
422
423 sched_setscheduler(current, policy, &param);
424}
425
426static void watchdog_enable(unsigned int cpu)
427{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200428 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400429
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200430 /*
431 * Start the timer first to prevent the NMI watchdog triggering
432 * before the timer has a chance to fire.
433 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100434 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
435 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800436 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400437 HRTIMER_MODE_REL_PINNED);
438
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200439 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000440 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200441 /* Enable the perf event */
442 watchdog_nmi_enable(cpu);
443
444 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400445}
446
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000447static void watchdog_disable(unsigned int cpu)
448{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200449 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000450
451 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200452 /*
453 * Disable the perf event first. That prevents that a large delay
454 * between disabling the timer and disabling the perf event causes
455 * the perf NMI to detect a false positive.
456 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000457 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200458 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000459}
460
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200461static void watchdog_cleanup(unsigned int cpu, bool online)
462{
463 watchdog_disable(cpu);
464}
465
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000466static int watchdog_should_run(unsigned int cpu)
467{
468 return __this_cpu_read(hrtimer_interrupts) !=
469 __this_cpu_read(soft_lockup_hrtimer_cnt);
470}
471
472/*
473 * The watchdog thread function - touches the timestamp.
474 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800475 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000476 * default) to reset the softlockup timestamp. If this gets delayed
477 * for more than 2*watchdog_thresh seconds then the debug-printout
478 * triggers in watchdog_timer_fn().
479 */
480static void watchdog(unsigned int cpu)
481{
482 __this_cpu_write(soft_lockup_hrtimer_cnt,
483 __this_cpu_read(hrtimer_interrupts));
484 __touch_watchdog();
485}
Don Zickus58687ac2010-05-07 17:11:44 -0400486
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200487static struct smp_hotplug_thread watchdog_threads = {
488 .store = &softlockup_watchdog,
489 .thread_should_run = watchdog_should_run,
490 .thread_fn = watchdog,
491 .thread_comm = "watchdog/%u",
492 .setup = watchdog_enable,
493 .cleanup = watchdog_cleanup,
494 .park = watchdog_disable,
495 .unpark = watchdog_enable,
496};
497
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200498static void softlockup_update_smpboot_threads(void)
499{
500 lockdep_assert_held(&watchdog_mutex);
501
502 if (!softlockup_threads_initialized)
503 return;
504
505 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
506 &watchdog_allowed_mask);
507 __lockup_detector_cleanup();
508}
509
510/* Temporarily park all watchdog threads */
511static void softlockup_park_all_threads(void)
512{
513 cpumask_clear(&watchdog_allowed_mask);
514 softlockup_update_smpboot_threads();
515}
516
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200517/* Unpark enabled threads */
518static void softlockup_unpark_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200519{
520 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
521 softlockup_update_smpboot_threads();
522}
523
524static void softlockup_reconfigure_threads(bool enabled)
525{
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200526 watchdog_nmi_reconfigure(false);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200527 softlockup_park_all_threads();
528 set_sample_period();
529 if (enabled)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200530 softlockup_unpark_threads();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200531 watchdog_nmi_reconfigure(true);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200532}
533
534/*
535 * Create the watchdog thread infrastructure.
536 *
537 * The threads are not unparked as watchdog_allowed_mask is empty. When
538 * the threads are sucessfully initialized, take the proper locks and
539 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
540 */
541static __init void softlockup_init_threads(void)
542{
543 int ret;
544
545 /*
546 * If sysctl is off and watchdog got disabled on the command line,
547 * nothing to do here.
548 */
549 if (!IS_ENABLED(CONFIG_SYSCTL) &&
550 !(watchdog_enabled && watchdog_thresh))
551 return;
552
553 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
554 &watchdog_allowed_mask);
555 if (ret) {
556 pr_err("Failed to initialize soft lockup detector threads\n");
557 return;
558 }
559
560 mutex_lock(&watchdog_mutex);
561 softlockup_threads_initialized = true;
562 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
563 mutex_unlock(&watchdog_mutex);
564}
565
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200566#else /* CONFIG_SOFTLOCKUP_DETECTOR */
567static inline int watchdog_park_threads(void) { return 0; }
568static inline void watchdog_unpark_threads(void) { }
569static inline int watchdog_enable_all_cpus(void) { return 0; }
570static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200571static inline void softlockup_init_threads(void) { }
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200572static void softlockup_reconfigure_threads(bool enabled)
573{
574 watchdog_nmi_reconfigure(false);
575 watchdog_nmi_reconfigure(true);
576}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200577#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700578
Thomas Gleixner941154b2017-09-12 21:37:04 +0200579static void __lockup_detector_cleanup(void)
580{
581 lockdep_assert_held(&watchdog_mutex);
582 hardlockup_detector_perf_cleanup();
583}
584
585/**
586 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
587 *
588 * Caller must not hold the cpu hotplug rwsem.
589 */
590void lockup_detector_cleanup(void)
591{
592 mutex_lock(&watchdog_mutex);
593 __lockup_detector_cleanup();
594 mutex_unlock(&watchdog_mutex);
595}
596
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200597/**
598 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
599 *
600 * Special interface for parisc. It prevents lockup detector warnings from
601 * the default pm_poweroff() function which busy loops forever.
602 */
603void lockup_detector_soft_poweroff(void)
604{
605 watchdog_enabled = 0;
606}
607
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800608#ifdef CONFIG_SYSCTL
609
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200610/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200611static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400612{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200613 /* Remove impossible cpus to keep sysctl output clean. */
614 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200615 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700616}
617
618/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700619 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
620 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200621 * caller | table->data points to | 'which'
622 * -------------------|----------------------------|--------------------------
623 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
624 * | | SOFT_WATCHDOG_ENABLED
625 * -------------------|----------------------------|--------------------------
626 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
627 * -------------------|----------------------------|--------------------------
628 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700629 */
630static int proc_watchdog_common(int which, struct ctl_table *table, int write,
631 void __user *buffer, size_t *lenp, loff_t *ppos)
632{
633 int err, old, new;
634 int *watchdog_param = (int *)table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400635
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200636 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200637 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000638
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700639 /*
640 * If the parameter is being read return the state of the corresponding
641 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
642 * run state of the lockup detectors.
643 */
644 if (!write) {
645 *watchdog_param = (watchdog_enabled & which) != 0;
646 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
647 } else {
648 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
649 if (err)
650 goto out;
651
652 /*
653 * There is a race window between fetching the current value
654 * from 'watchdog_enabled' and storing the new value. During
655 * this race window, watchdog_nmi_enable() can sneak in and
656 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
657 * The 'cmpxchg' detects this race and the loop retries.
658 */
659 do {
660 old = watchdog_enabled;
661 /*
662 * If the parameter value is not zero set the
663 * corresponding bit(s), else clear it(them).
664 */
665 if (*watchdog_param)
666 new = old | which;
667 else
668 new = old & ~which;
669 } while (cmpxchg(&watchdog_enabled, old, new) != old);
670
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200671 if (old != new)
672 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700673 }
674out:
Thomas Gleixner946d1972017-09-12 21:37:01 +0200675 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200676 cpu_hotplug_enable();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700677 return err;
678}
679
680/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700681 * /proc/sys/kernel/watchdog
682 */
683int proc_watchdog(struct ctl_table *table, int write,
684 void __user *buffer, size_t *lenp, loff_t *ppos)
685{
686 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
687 table, write, buffer, lenp, ppos);
688}
689
690/*
691 * /proc/sys/kernel/nmi_watchdog
692 */
693int proc_nmi_watchdog(struct ctl_table *table, int write,
694 void __user *buffer, size_t *lenp, loff_t *ppos)
695{
696 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
697 table, write, buffer, lenp, ppos);
698}
699
700/*
701 * /proc/sys/kernel/soft_watchdog
702 */
703int proc_soft_watchdog(struct ctl_table *table, int write,
704 void __user *buffer, size_t *lenp, loff_t *ppos)
705{
706 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
707 table, write, buffer, lenp, ppos);
708}
709
710/*
711 * /proc/sys/kernel/watchdog_thresh
712 */
713int proc_watchdog_thresh(struct ctl_table *table, int write,
714 void __user *buffer, size_t *lenp, loff_t *ppos)
715{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200716 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700717
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200718 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200719 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700720
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200721 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200722 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700723
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200724 if (!err && write && old != READ_ONCE(watchdog_thresh))
725 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700726
Thomas Gleixner946d1972017-09-12 21:37:01 +0200727 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200728 cpu_hotplug_enable();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200729 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400730}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700731
732/*
733 * The cpumask is the mask of possible cpus that the watchdog can run
734 * on, not the mask of cpus it is actually running on. This allows the
735 * user to specify a mask that will include cpus that have not yet
736 * been brought online, if desired.
737 */
738int proc_watchdog_cpumask(struct ctl_table *table, int write,
739 void __user *buffer, size_t *lenp, loff_t *ppos)
740{
741 int err;
742
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200743 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200744 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700745
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700746 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200747 if (!err && write)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200748 proc_watchdog_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200749
Thomas Gleixner946d1972017-09-12 21:37:01 +0200750 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200751 cpu_hotplug_enable();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700752 return err;
753}
Don Zickus58687ac2010-05-07 17:11:44 -0400754#endif /* CONFIG_SYSCTL */
755
Peter Zijlstra004417a2010-11-25 18:38:29 +0100756void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400757{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700758#ifdef CONFIG_NO_HZ_FULL
759 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700760 pr_info("Disabling watchdog on nohz_full cores by default\n");
761 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700762 } else
763 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
764#else
765 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
766#endif
767
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200768 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400769}