blob: 762d3ed82a087ac0e341e647c62d3e57ed695e55 [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 -070034int __read_mostly nmi_watchdog_enabled;
35
36#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
37unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED |
38 NMI_WATCHDOG_ENABLED;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070039#else
Babu Moger249e52e2016-12-14 15:06:21 -080040unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070041#endif
Nicholas Piggin05a4a952017-07-12 14:35:46 -070042
43#ifdef CONFIG_HARDLOCKUP_DETECTOR
Nicholas Piggin05a4a952017-07-12 14:35:46 -070044/*
45 * Should we panic when a soft-lockup or hard-lockup occurs:
46 */
47unsigned int __read_mostly hardlockup_panic =
48 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
49/*
50 * We may not want to enable hard lockup detection by default in all cases,
51 * for example when running the kernel as a guest on a hypervisor. In these
52 * cases this function can be called to disable hard lockup detection. This
53 * function should only be executed once by the boot processor before the
54 * kernel command line parameters are parsed, because otherwise it is not
55 * possible to override this in hardlockup_panic_setup().
56 */
Thomas Gleixner7a355822017-09-12 21:37:02 +020057void __init hardlockup_detector_disable(void)
Nicholas Piggin05a4a952017-07-12 14:35:46 -070058{
59 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
60}
61
62static int __init hardlockup_panic_setup(char *str)
63{
64 if (!strncmp(str, "panic", 5))
65 hardlockup_panic = 1;
66 else if (!strncmp(str, "nopanic", 7))
67 hardlockup_panic = 0;
68 else if (!strncmp(str, "0", 1))
69 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
70 else if (!strncmp(str, "1", 1))
71 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
72 return 1;
73}
74__setup("nmi_watchdog=", hardlockup_panic_setup);
75
Thomas Gleixner368a7e22017-09-12 21:37:07 +020076# ifdef CONFIG_SMP
77int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070078
Thomas Gleixner368a7e22017-09-12 21:37:07 +020079static int __init hardlockup_all_cpu_backtrace_setup(char *str)
80{
81 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
82 return 1;
83}
84__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
85# endif /* CONFIG_SMP */
86#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -070087
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070088int __read_mostly watchdog_user_enabled;
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -070089int __read_mostly watchdog_thresh = 10;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070090
Nicholas Piggin05a4a952017-07-12 14:35:46 -070091struct cpumask watchdog_cpumask __read_mostly;
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070092unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
93
Ulrich Obergfellec6a9062015-09-04 15:45:28 -070094/*
Nicholas Piggin05a4a952017-07-12 14:35:46 -070095 * These functions can be overridden if an architecture implements its
96 * own hardlockup detector.
Nicholas Piggina10a8422017-07-12 14:35:49 -070097 *
98 * watchdog_nmi_enable/disable can be implemented to start and stop when
99 * softlockup watchdog threads start and stop. The arch must select the
100 * SOFTLOCKUP_DETECTOR Kconfig.
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700101 */
102int __weak watchdog_nmi_enable(unsigned int cpu)
103{
104 return 0;
105}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200106
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700107void __weak watchdog_nmi_disable(unsigned int cpu)
108{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200109 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700110}
111
Nicholas Piggina10a8422017-07-12 14:35:49 -0700112/*
113 * watchdog_nmi_reconfigure can be implemented to be notified after any
114 * watchdog configuration change. The arch hardlockup watchdog should
115 * respond to the following variables:
116 * - nmi_watchdog_enabled
117 * - watchdog_thresh
118 * - watchdog_cpumask
119 * - sysctl_hardlockup_all_cpu_backtrace
120 * - hardlockup_panic
Nicholas Piggina10a8422017-07-12 14:35:49 -0700121 */
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200122void __weak watchdog_nmi_reconfigure(void) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700123
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700124#ifdef CONFIG_SOFTLOCKUP_DETECTOR
125
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200126/* Global variables, exported for sysctl */
127unsigned int __read_mostly softlockup_panic =
128 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
129int __read_mostly soft_watchdog_enabled;
130
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200131struct cpumask watchdog_allowed_mask __read_mostly;
132static bool softlockup_threads_initialized __read_mostly;
133
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800134static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400135
136static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
137static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
138static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
139static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400140static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000141static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
142static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700143static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400144static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700145static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400146
Don Zickus58687ac2010-05-07 17:11:44 -0400147static int __init softlockup_panic_setup(char *str)
148{
149 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400150 return 1;
151}
152__setup("softlockup_panic=", softlockup_panic_setup);
153
154static int __init nowatchdog_setup(char *str)
155{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700156 watchdog_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400157 return 1;
158}
159__setup("nowatchdog", nowatchdog_setup);
160
Don Zickus58687ac2010-05-07 17:11:44 -0400161static int __init nosoftlockup_setup(char *str)
162{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700163 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
Don Zickus58687ac2010-05-07 17:11:44 -0400164 return 1;
165}
166__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700167
Aaron Tomlined235872014-06-23 13:22:05 -0700168#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200169int __read_mostly sysctl_softlockup_all_cpu_backtrace;
170
Aaron Tomlined235872014-06-23 13:22:05 -0700171static int __init softlockup_all_cpu_backtrace_setup(char *str)
172{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200173 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700174 return 1;
175}
176__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700177#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400178
Thomas Gleixner941154b2017-09-12 21:37:04 +0200179static void __lockup_detector_cleanup(void);
180
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700181/*
182 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
183 * lockups can have false positives under extreme conditions. So we generally
184 * want a higher threshold for soft lockups than for hard lockups. So we couple
185 * the thresholds with a factor: we make the soft threshold twice the amount of
186 * time the hard threshold is.
187 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200188static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700189{
190 return watchdog_thresh * 2;
191}
Don Zickus58687ac2010-05-07 17:11:44 -0400192
193/*
194 * Returns seconds, approximately. We don't need nanosecond
195 * resolution, and we don't need to waste time with a big divide when
196 * 2^30ns == 1.074s.
197 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900198static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400199{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800200 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400201}
202
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800203static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400204{
205 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700206 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500207 * the divide by 5 is to give hrtimer several chances (two
208 * or three with the current relation between the soft
209 * and hard thresholds) to increment before the
210 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400211 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800212 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200213 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400214}
215
216/* Commands for resetting the watchdog */
217static void __touch_watchdog(void)
218{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900219 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400220}
221
Tejun Heo03e0d462015-12-08 11:28:04 -0500222/**
223 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
224 *
225 * Call when the scheduler may have stalled for legitimate reasons
226 * preventing the watchdog task from executing - e.g. the scheduler
227 * entering idle state. This should only be used for scheduler events.
228 * Use touch_softlockup_watchdog() for everything else.
229 */
230void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400231{
Andrew Morton78611442014-04-18 15:07:12 -0700232 /*
233 * Preemption can be enabled. It doesn't matter which CPU's timestamp
234 * gets zeroed here, so use the raw_ operation.
235 */
236 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400237}
Tejun Heo03e0d462015-12-08 11:28:04 -0500238
239void touch_softlockup_watchdog(void)
240{
241 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500242 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500243}
Ingo Molnar0167c782010-05-13 08:53:33 +0200244EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400245
Don Zickus332fbdb2010-05-07 17:11:45 -0400246void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400247{
248 int cpu;
249
250 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200251 * watchdog_mutex cannpt be taken here, as this might be called
252 * from (soft)interrupt context, so the access to
253 * watchdog_allowed_cpumask might race with a concurrent update.
254 *
255 * The watchdog time stamp can race against a concurrent real
256 * update as well, the only side effect might be a cycle delay for
257 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400258 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200259 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400260 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500261 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400262}
263
Don Zickus58687ac2010-05-07 17:11:44 -0400264void touch_softlockup_watchdog_sync(void)
265{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500266 __this_cpu_write(softlockup_touch_sync, true);
267 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400268}
269
Don Zickus26e09c62010-05-17 18:06:04 -0400270static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400271{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900272 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400273
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800274 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700275 /* Warn about unreasonable delays. */
276 if (time_after(now, touch_ts + get_softlockup_thresh()))
277 return now - touch_ts;
278 }
Don Zickus58687ac2010-05-07 17:11:44 -0400279 return 0;
280}
281
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700282/* watchdog detector functions */
283bool is_hardlockup(void)
284{
285 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
286
287 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
288 return true;
289
290 __this_cpu_write(hrtimer_interrupts_saved, hrint);
291 return false;
292}
293
Don Zickus58687ac2010-05-07 17:11:44 -0400294static void watchdog_interrupt_count(void)
295{
Christoph Lameter909ea962010-12-08 16:22:55 +0100296 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400297}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000298
Don Zickus58687ac2010-05-07 17:11:44 -0400299/* watchdog kicker functions */
300static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
301{
Christoph Lameter909ea962010-12-08 16:22:55 +0100302 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400303 struct pt_regs *regs = get_irq_regs();
304 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700305 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400306
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200307 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800308 return HRTIMER_NORESTART;
309
Don Zickus58687ac2010-05-07 17:11:44 -0400310 /* kick the hardlockup detector */
311 watchdog_interrupt_count();
312
313 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100314 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400315
316 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800317 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400318
319 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100320 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400321 /*
322 * If the time stamp was touched atomically
323 * make sure the scheduler tick is up to date.
324 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100325 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400326 sched_clock_tick();
327 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500328
329 /* Clear the guest paused flag on watchdog reset */
330 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400331 __touch_watchdog();
332 return HRTIMER_RESTART;
333 }
334
335 /* check for a softlockup
336 * This is done by making sure a high priority task is
337 * being scheduled. The task touches the watchdog to
338 * indicate it is getting cpu time. If it hasn't then
339 * this is a good indication some task is hogging the cpu
340 */
Don Zickus26e09c62010-05-17 18:06:04 -0400341 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400342 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500343 /*
344 * If a virtual machine is stopped by the host it can look to
345 * the watchdog like a soft lockup, check to see if the host
346 * stopped the vm before we issue the warning
347 */
348 if (kvm_check_and_clear_guest_paused())
349 return HRTIMER_RESTART;
350
Don Zickus58687ac2010-05-07 17:11:44 -0400351 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700352 if (__this_cpu_read(soft_watchdog_warn) == true) {
353 /*
354 * When multiple processes are causing softlockups the
355 * softlockup detector only warns on the first one
356 * because the code relies on a full quiet cycle to
357 * re-arm. The second process prevents the quiet cycle
358 * and never gets reported. Use task pointers to detect
359 * this.
360 */
361 if (__this_cpu_read(softlockup_task_ptr_saved) !=
362 current) {
363 __this_cpu_write(soft_watchdog_warn, false);
364 __touch_watchdog();
365 }
Don Zickus58687ac2010-05-07 17:11:44 -0400366 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700367 }
Don Zickus58687ac2010-05-07 17:11:44 -0400368
Aaron Tomlined235872014-06-23 13:22:05 -0700369 if (softlockup_all_cpu_backtrace) {
370 /* Prevent multiple soft-lockup reports if one cpu is already
371 * engaged in dumping cpu back traces
372 */
373 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
374 /* Someone else will report us. Let's give up */
375 __this_cpu_write(soft_watchdog_warn, true);
376 return HRTIMER_RESTART;
377 }
378 }
379
Fabian Frederick656c3b72014-08-06 16:04:03 -0700380 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400381 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400382 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700383 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400384 print_modules();
385 print_irqtrace_events(current);
386 if (regs)
387 show_regs(regs);
388 else
389 dump_stack();
390
Aaron Tomlined235872014-06-23 13:22:05 -0700391 if (softlockup_all_cpu_backtrace) {
392 /* Avoid generating two back traces for current
393 * given that one is already made above
394 */
395 trigger_allbutself_cpu_backtrace();
396
397 clear_bit(0, &soft_lockup_nmi_warn);
398 /* Barrier to sync with other cpus */
399 smp_mb__after_atomic();
400 }
401
Josh Hunt69361ee2014-08-08 14:22:31 -0700402 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400403 if (softlockup_panic)
404 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100405 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400406 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100407 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400408
409 return HRTIMER_RESTART;
410}
411
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000412static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400413{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000414 struct sched_param param = { .sched_priority = prio };
415
416 sched_setscheduler(current, policy, &param);
417}
418
419static void watchdog_enable(unsigned int cpu)
420{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200421 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400422
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200423 /*
424 * Start the timer first to prevent the NMI watchdog triggering
425 * before the timer has a chance to fire.
426 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100427 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
428 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800429 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400430 HRTIMER_MODE_REL_PINNED);
431
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200432 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000433 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200434 /* Enable the perf event */
435 watchdog_nmi_enable(cpu);
436
437 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400438}
439
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000440static void watchdog_disable(unsigned int cpu)
441{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200442 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000443
444 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200445 /*
446 * Disable the perf event first. That prevents that a large delay
447 * between disabling the timer and disabling the perf event causes
448 * the perf NMI to detect a false positive.
449 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000450 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200451 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000452}
453
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200454static void watchdog_cleanup(unsigned int cpu, bool online)
455{
456 watchdog_disable(cpu);
457}
458
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000459static int watchdog_should_run(unsigned int cpu)
460{
461 return __this_cpu_read(hrtimer_interrupts) !=
462 __this_cpu_read(soft_lockup_hrtimer_cnt);
463}
464
465/*
466 * The watchdog thread function - touches the timestamp.
467 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800468 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000469 * default) to reset the softlockup timestamp. If this gets delayed
470 * for more than 2*watchdog_thresh seconds then the debug-printout
471 * triggers in watchdog_timer_fn().
472 */
473static void watchdog(unsigned int cpu)
474{
475 __this_cpu_write(soft_lockup_hrtimer_cnt,
476 __this_cpu_read(hrtimer_interrupts));
477 __touch_watchdog();
478}
Don Zickus58687ac2010-05-07 17:11:44 -0400479
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200480static struct smp_hotplug_thread watchdog_threads = {
481 .store = &softlockup_watchdog,
482 .thread_should_run = watchdog_should_run,
483 .thread_fn = watchdog,
484 .thread_comm = "watchdog/%u",
485 .setup = watchdog_enable,
486 .cleanup = watchdog_cleanup,
487 .park = watchdog_disable,
488 .unpark = watchdog_enable,
489};
490
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200491static void softlockup_update_smpboot_threads(void)
492{
493 lockdep_assert_held(&watchdog_mutex);
494
495 if (!softlockup_threads_initialized)
496 return;
497
498 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
499 &watchdog_allowed_mask);
500 __lockup_detector_cleanup();
501}
502
503/* Temporarily park all watchdog threads */
504static void softlockup_park_all_threads(void)
505{
506 cpumask_clear(&watchdog_allowed_mask);
507 softlockup_update_smpboot_threads();
508}
509
510/*
511 * Park threads which are not longer enabled and unpark threads which have
512 * been newly enabled.
513 */
514static void softlockup_update_threads(void)
515{
516 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
517 softlockup_update_smpboot_threads();
518}
519
520static void softlockup_reconfigure_threads(bool enabled)
521{
522 softlockup_park_all_threads();
523 set_sample_period();
524 if (enabled)
525 softlockup_update_threads();
526}
527
528/*
529 * Create the watchdog thread infrastructure.
530 *
531 * The threads are not unparked as watchdog_allowed_mask is empty. When
532 * the threads are sucessfully initialized, take the proper locks and
533 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
534 */
535static __init void softlockup_init_threads(void)
536{
537 int ret;
538
539 /*
540 * If sysctl is off and watchdog got disabled on the command line,
541 * nothing to do here.
542 */
543 if (!IS_ENABLED(CONFIG_SYSCTL) &&
544 !(watchdog_enabled && watchdog_thresh))
545 return;
546
547 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
548 &watchdog_allowed_mask);
549 if (ret) {
550 pr_err("Failed to initialize soft lockup detector threads\n");
551 return;
552 }
553
554 mutex_lock(&watchdog_mutex);
555 softlockup_threads_initialized = true;
556 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
557 mutex_unlock(&watchdog_mutex);
558}
559
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200560#else /* CONFIG_SOFTLOCKUP_DETECTOR */
561static inline int watchdog_park_threads(void) { return 0; }
562static inline void watchdog_unpark_threads(void) { }
563static inline int watchdog_enable_all_cpus(void) { return 0; }
564static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200565static inline void softlockup_init_threads(void) { }
566static inline void softlockup_update_threads(void) { }
567static inline void softlockup_reconfigure_threads(bool enabled) { }
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200568#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700569
Thomas Gleixner941154b2017-09-12 21:37:04 +0200570static void __lockup_detector_cleanup(void)
571{
572 lockdep_assert_held(&watchdog_mutex);
573 hardlockup_detector_perf_cleanup();
574}
575
576/**
577 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
578 *
579 * Caller must not hold the cpu hotplug rwsem.
580 */
581void lockup_detector_cleanup(void)
582{
583 mutex_lock(&watchdog_mutex);
584 __lockup_detector_cleanup();
585 mutex_unlock(&watchdog_mutex);
586}
587
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200588/**
589 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
590 *
591 * Special interface for parisc. It prevents lockup detector warnings from
592 * the default pm_poweroff() function which busy loops forever.
593 */
594void lockup_detector_soft_poweroff(void)
595{
596 watchdog_enabled = 0;
597}
598
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800599#ifdef CONFIG_SYSCTL
600
Don Zickus58687ac2010-05-07 17:11:44 -0400601/*
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700602 * Update the run state of the lockup detectors.
Don Zickus58687ac2010-05-07 17:11:44 -0400603 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200604static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400605{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200606 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
Nicholas Piggina10a8422017-07-12 14:35:49 -0700607 watchdog_nmi_reconfigure();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700608}
609
610/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700611 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
612 *
613 * caller | table->data points to | 'which' contains the flag(s)
614 * -------------------|-----------------------|-----------------------------
615 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
616 * | | with SOFT_WATCHDOG_ENABLED
617 * -------------------|-----------------------|-----------------------------
618 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
619 * -------------------|-----------------------|-----------------------------
620 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
621 */
622static int proc_watchdog_common(int which, struct ctl_table *table, int write,
623 void __user *buffer, size_t *lenp, loff_t *ppos)
624{
625 int err, old, new;
626 int *watchdog_param = (int *)table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400627
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200628 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200629 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000630
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700631 /*
632 * If the parameter is being read return the state of the corresponding
633 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
634 * run state of the lockup detectors.
635 */
636 if (!write) {
637 *watchdog_param = (watchdog_enabled & which) != 0;
638 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
639 } else {
640 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
641 if (err)
642 goto out;
643
644 /*
645 * There is a race window between fetching the current value
646 * from 'watchdog_enabled' and storing the new value. During
647 * this race window, watchdog_nmi_enable() can sneak in and
648 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
649 * The 'cmpxchg' detects this race and the loop retries.
650 */
651 do {
652 old = watchdog_enabled;
653 /*
654 * If the parameter value is not zero set the
655 * corresponding bit(s), else clear it(them).
656 */
657 if (*watchdog_param)
658 new = old | which;
659 else
660 new = old & ~which;
661 } while (cmpxchg(&watchdog_enabled, old, new) != old);
662
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200663 if (old != new)
664 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700665 }
666out:
Thomas Gleixner946d1972017-09-12 21:37:01 +0200667 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200668 cpu_hotplug_enable();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700669 return err;
670}
671
672/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700673 * /proc/sys/kernel/watchdog
674 */
675int proc_watchdog(struct ctl_table *table, int write,
676 void __user *buffer, size_t *lenp, loff_t *ppos)
677{
678 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
679 table, write, buffer, lenp, ppos);
680}
681
682/*
683 * /proc/sys/kernel/nmi_watchdog
684 */
685int proc_nmi_watchdog(struct ctl_table *table, int write,
686 void __user *buffer, size_t *lenp, loff_t *ppos)
687{
688 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
689 table, write, buffer, lenp, ppos);
690}
691
692/*
693 * /proc/sys/kernel/soft_watchdog
694 */
695int proc_soft_watchdog(struct ctl_table *table, int write,
696 void __user *buffer, size_t *lenp, loff_t *ppos)
697{
698 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
699 table, write, buffer, lenp, ppos);
700}
701
702/*
703 * /proc/sys/kernel/watchdog_thresh
704 */
705int proc_watchdog_thresh(struct ctl_table *table, int write,
706 void __user *buffer, size_t *lenp, loff_t *ppos)
707{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200708 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700709
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200710 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200711 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700712
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200713 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200714 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700715
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200716 if (!err && write && old != READ_ONCE(watchdog_thresh))
717 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700718
Thomas Gleixner946d1972017-09-12 21:37:01 +0200719 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200720 cpu_hotplug_enable();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200721 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400722}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700723
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200724static void proc_watchdog_cpumask_update(void)
725{
726 /* Remove impossible cpus to keep sysctl output clean. */
727 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200728
729 softlockup_update_threads();
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200730 watchdog_nmi_reconfigure();
731}
732
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700733/*
734 * The cpumask is the mask of possible cpus that the watchdog can run
735 * on, not the mask of cpus it is actually running on. This allows the
736 * user to specify a mask that will include cpus that have not yet
737 * been brought online, if desired.
738 */
739int proc_watchdog_cpumask(struct ctl_table *table, int write,
740 void __user *buffer, size_t *lenp, loff_t *ppos)
741{
742 int err;
743
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200744 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200745 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700746
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700747 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200748 if (!err && write)
749 proc_watchdog_cpumask_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200750
Thomas Gleixner946d1972017-09-12 21:37:01 +0200751 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200752 cpu_hotplug_enable();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700753 return err;
754}
Don Zickus58687ac2010-05-07 17:11:44 -0400755#endif /* CONFIG_SYSCTL */
756
Peter Zijlstra004417a2010-11-25 18:38:29 +0100757void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400758{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700759#ifdef CONFIG_NO_HZ_FULL
760 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700761 pr_info("Disabling watchdog on nohz_full cores by default\n");
762 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700763 } else
764 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
765#else
766 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
767#endif
768
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200769 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400770}