blob: 5eb11960e4a26a4bdd371e289b5f87cc1d10c1ae [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)
Thomas Gleixner09154982017-09-12 21:37:17 +020035# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
36# define NMI_WATCHDOG_DEFAULT 1
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070037#else
Thomas Gleixner09154982017-09-12 21:37:17 +020038# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
39# define NMI_WATCHDOG_DEFAULT 0
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070040#endif
Nicholas Piggin05a4a952017-07-12 14:35:46 -070041
Thomas Gleixner09154982017-09-12 21:37:17 +020042unsigned long __read_mostly watchdog_enabled;
43int __read_mostly watchdog_user_enabled = 1;
44int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
45int __read_mostly soft_watchdog_user_enabled = 1;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020046int __read_mostly watchdog_thresh = 10;
Thomas Gleixnera994a312017-09-12 21:37:19 +020047int __read_mostly nmi_watchdog_available;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020048
49struct cpumask watchdog_allowed_mask __read_mostly;
50static bool softlockup_threads_initialized __read_mostly;
51
52struct cpumask watchdog_cpumask __read_mostly;
53unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
54
Nicholas Piggin05a4a952017-07-12 14:35:46 -070055#ifdef CONFIG_HARDLOCKUP_DETECTOR
Nicholas Piggin05a4a952017-07-12 14:35:46 -070056/*
57 * Should we panic when a soft-lockup or hard-lockup occurs:
58 */
59unsigned int __read_mostly hardlockup_panic =
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
61/*
62 * We may not want to enable hard lockup detection by default in all cases,
63 * for example when running the kernel as a guest on a hypervisor. In these
64 * cases this function can be called to disable hard lockup detection. This
65 * function should only be executed once by the boot processor before the
66 * kernel command line parameters are parsed, because otherwise it is not
67 * possible to override this in hardlockup_panic_setup().
68 */
Thomas Gleixner7a355822017-09-12 21:37:02 +020069void __init hardlockup_detector_disable(void)
Nicholas Piggin05a4a952017-07-12 14:35:46 -070070{
Thomas Gleixner09154982017-09-12 21:37:17 +020071 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070072}
73
74static int __init hardlockup_panic_setup(char *str)
75{
76 if (!strncmp(str, "panic", 5))
77 hardlockup_panic = 1;
78 else if (!strncmp(str, "nopanic", 7))
79 hardlockup_panic = 0;
80 else if (!strncmp(str, "0", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020081 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070082 else if (!strncmp(str, "1", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020083 nmi_watchdog_user_enabled = 1;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070084 return 1;
85}
86__setup("nmi_watchdog=", hardlockup_panic_setup);
87
Thomas Gleixner368a7e22017-09-12 21:37:07 +020088# ifdef CONFIG_SMP
89int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070090
Thomas Gleixner368a7e22017-09-12 21:37:07 +020091static int __init hardlockup_all_cpu_backtrace_setup(char *str)
92{
93 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
94 return 1;
95}
96__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
97# endif /* CONFIG_SMP */
98#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -070099
Ulrich Obergfellec6a9062015-09-04 15:45:28 -0700100/*
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700101 * These functions can be overridden if an architecture implements its
102 * own hardlockup detector.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700103 *
104 * watchdog_nmi_enable/disable can be implemented to start and stop when
105 * softlockup watchdog threads start and stop. The arch must select the
106 * SOFTLOCKUP_DETECTOR Kconfig.
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700107 */
108int __weak watchdog_nmi_enable(unsigned int cpu)
109{
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200110 hardlockup_detector_perf_enable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700111 return 0;
112}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200113
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700114void __weak watchdog_nmi_disable(unsigned int cpu)
115{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200116 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700117}
118
Thomas Gleixnera994a312017-09-12 21:37:19 +0200119/* Return 0, if a NMI watchdog is available. Error code otherwise */
120int __weak __init watchdog_nmi_probe(void)
121{
122 return hardlockup_detector_perf_init();
123}
124
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200125/**
126 * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs
127 * @run: If false stop the watchdogs on all enabled CPUs
128 * If true start the watchdogs on all enabled CPUs
129 *
130 * The core call order is:
131 * watchdog_nmi_reconfigure(false);
132 * update_variables();
133 * watchdog_nmi_reconfigure(true);
134 *
135 * The second call which starts the watchdogs again guarantees that the
136 * following variables are stable across the call.
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200137 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700138 * - watchdog_thresh
139 * - watchdog_cpumask
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200140 *
141 * After the call the variables can be changed again.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700142 */
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200143void __weak watchdog_nmi_reconfigure(bool run) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700144
Thomas Gleixner09154982017-09-12 21:37:17 +0200145/**
146 * lockup_detector_update_enable - Update the sysctl enable bit
147 *
148 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
149 * can't race with watchdog_nmi_disable().
150 */
151static void lockup_detector_update_enable(void)
152{
153 watchdog_enabled = 0;
154 if (!watchdog_user_enabled)
155 return;
Thomas Gleixnera994a312017-09-12 21:37:19 +0200156 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
Thomas Gleixner09154982017-09-12 21:37:17 +0200157 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
158 if (soft_watchdog_user_enabled)
159 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
160}
161
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700162#ifdef CONFIG_SOFTLOCKUP_DETECTOR
163
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200164/* Global variables, exported for sysctl */
165unsigned int __read_mostly softlockup_panic =
166 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200167
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800168static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400169
170static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
171static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
172static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
173static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400174static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000175static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
176static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700177static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400178static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700179static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400180
Don Zickus58687ac2010-05-07 17:11:44 -0400181static int __init softlockup_panic_setup(char *str)
182{
183 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400184 return 1;
185}
186__setup("softlockup_panic=", softlockup_panic_setup);
187
188static int __init nowatchdog_setup(char *str)
189{
Thomas Gleixner09154982017-09-12 21:37:17 +0200190 watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400191 return 1;
192}
193__setup("nowatchdog", nowatchdog_setup);
194
Don Zickus58687ac2010-05-07 17:11:44 -0400195static int __init nosoftlockup_setup(char *str)
196{
Thomas Gleixner09154982017-09-12 21:37:17 +0200197 soft_watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400198 return 1;
199}
200__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700201
Aaron Tomlined235872014-06-23 13:22:05 -0700202#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200203int __read_mostly sysctl_softlockup_all_cpu_backtrace;
204
Aaron Tomlined235872014-06-23 13:22:05 -0700205static int __init softlockup_all_cpu_backtrace_setup(char *str)
206{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200207 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700208 return 1;
209}
210__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700211#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400212
Thomas Gleixner941154b2017-09-12 21:37:04 +0200213static void __lockup_detector_cleanup(void);
214
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700215/*
216 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
217 * lockups can have false positives under extreme conditions. So we generally
218 * want a higher threshold for soft lockups than for hard lockups. So we couple
219 * the thresholds with a factor: we make the soft threshold twice the amount of
220 * time the hard threshold is.
221 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200222static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700223{
224 return watchdog_thresh * 2;
225}
Don Zickus58687ac2010-05-07 17:11:44 -0400226
227/*
228 * Returns seconds, approximately. We don't need nanosecond
229 * resolution, and we don't need to waste time with a big divide when
230 * 2^30ns == 1.074s.
231 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900232static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400233{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800234 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400235}
236
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800237static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400238{
239 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700240 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500241 * the divide by 5 is to give hrtimer several chances (two
242 * or three with the current relation between the soft
243 * and hard thresholds) to increment before the
244 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400245 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800246 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200247 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400248}
249
250/* Commands for resetting the watchdog */
251static void __touch_watchdog(void)
252{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900253 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400254}
255
Tejun Heo03e0d462015-12-08 11:28:04 -0500256/**
257 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
258 *
259 * Call when the scheduler may have stalled for legitimate reasons
260 * preventing the watchdog task from executing - e.g. the scheduler
261 * entering idle state. This should only be used for scheduler events.
262 * Use touch_softlockup_watchdog() for everything else.
263 */
264void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400265{
Andrew Morton78611442014-04-18 15:07:12 -0700266 /*
267 * Preemption can be enabled. It doesn't matter which CPU's timestamp
268 * gets zeroed here, so use the raw_ operation.
269 */
270 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400271}
Tejun Heo03e0d462015-12-08 11:28:04 -0500272
273void touch_softlockup_watchdog(void)
274{
275 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500276 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500277}
Ingo Molnar0167c782010-05-13 08:53:33 +0200278EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400279
Don Zickus332fbdb2010-05-07 17:11:45 -0400280void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400281{
282 int cpu;
283
284 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200285 * watchdog_mutex cannpt be taken here, as this might be called
286 * from (soft)interrupt context, so the access to
287 * watchdog_allowed_cpumask might race with a concurrent update.
288 *
289 * The watchdog time stamp can race against a concurrent real
290 * update as well, the only side effect might be a cycle delay for
291 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400292 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200293 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400294 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500295 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400296}
297
Don Zickus58687ac2010-05-07 17:11:44 -0400298void touch_softlockup_watchdog_sync(void)
299{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500300 __this_cpu_write(softlockup_touch_sync, true);
301 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400302}
303
Don Zickus26e09c62010-05-17 18:06:04 -0400304static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400305{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900306 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400307
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800308 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700309 /* Warn about unreasonable delays. */
310 if (time_after(now, touch_ts + get_softlockup_thresh()))
311 return now - touch_ts;
312 }
Don Zickus58687ac2010-05-07 17:11:44 -0400313 return 0;
314}
315
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700316/* watchdog detector functions */
317bool is_hardlockup(void)
318{
319 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
320
321 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
322 return true;
323
324 __this_cpu_write(hrtimer_interrupts_saved, hrint);
325 return false;
326}
327
Don Zickus58687ac2010-05-07 17:11:44 -0400328static void watchdog_interrupt_count(void)
329{
Christoph Lameter909ea962010-12-08 16:22:55 +0100330 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400331}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000332
Don Zickus58687ac2010-05-07 17:11:44 -0400333/* watchdog kicker functions */
334static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
335{
Christoph Lameter909ea962010-12-08 16:22:55 +0100336 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400337 struct pt_regs *regs = get_irq_regs();
338 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700339 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400340
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200341 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800342 return HRTIMER_NORESTART;
343
Don Zickus58687ac2010-05-07 17:11:44 -0400344 /* kick the hardlockup detector */
345 watchdog_interrupt_count();
346
347 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100348 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400349
350 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800351 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400352
353 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100354 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400355 /*
356 * If the time stamp was touched atomically
357 * make sure the scheduler tick is up to date.
358 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100359 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400360 sched_clock_tick();
361 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500362
363 /* Clear the guest paused flag on watchdog reset */
364 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400365 __touch_watchdog();
366 return HRTIMER_RESTART;
367 }
368
369 /* check for a softlockup
370 * This is done by making sure a high priority task is
371 * being scheduled. The task touches the watchdog to
372 * indicate it is getting cpu time. If it hasn't then
373 * this is a good indication some task is hogging the cpu
374 */
Don Zickus26e09c62010-05-17 18:06:04 -0400375 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400376 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500377 /*
378 * If a virtual machine is stopped by the host it can look to
379 * the watchdog like a soft lockup, check to see if the host
380 * stopped the vm before we issue the warning
381 */
382 if (kvm_check_and_clear_guest_paused())
383 return HRTIMER_RESTART;
384
Don Zickus58687ac2010-05-07 17:11:44 -0400385 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700386 if (__this_cpu_read(soft_watchdog_warn) == true) {
387 /*
388 * When multiple processes are causing softlockups the
389 * softlockup detector only warns on the first one
390 * because the code relies on a full quiet cycle to
391 * re-arm. The second process prevents the quiet cycle
392 * and never gets reported. Use task pointers to detect
393 * this.
394 */
395 if (__this_cpu_read(softlockup_task_ptr_saved) !=
396 current) {
397 __this_cpu_write(soft_watchdog_warn, false);
398 __touch_watchdog();
399 }
Don Zickus58687ac2010-05-07 17:11:44 -0400400 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700401 }
Don Zickus58687ac2010-05-07 17:11:44 -0400402
Aaron Tomlined235872014-06-23 13:22:05 -0700403 if (softlockup_all_cpu_backtrace) {
404 /* Prevent multiple soft-lockup reports if one cpu is already
405 * engaged in dumping cpu back traces
406 */
407 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
408 /* Someone else will report us. Let's give up */
409 __this_cpu_write(soft_watchdog_warn, true);
410 return HRTIMER_RESTART;
411 }
412 }
413
Fabian Frederick656c3b72014-08-06 16:04:03 -0700414 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400415 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400416 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700417 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400418 print_modules();
419 print_irqtrace_events(current);
420 if (regs)
421 show_regs(regs);
422 else
423 dump_stack();
424
Aaron Tomlined235872014-06-23 13:22:05 -0700425 if (softlockup_all_cpu_backtrace) {
426 /* Avoid generating two back traces for current
427 * given that one is already made above
428 */
429 trigger_allbutself_cpu_backtrace();
430
431 clear_bit(0, &soft_lockup_nmi_warn);
432 /* Barrier to sync with other cpus */
433 smp_mb__after_atomic();
434 }
435
Josh Hunt69361ee2014-08-08 14:22:31 -0700436 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400437 if (softlockup_panic)
438 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100439 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400440 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100441 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400442
443 return HRTIMER_RESTART;
444}
445
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000446static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400447{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000448 struct sched_param param = { .sched_priority = prio };
449
450 sched_setscheduler(current, policy, &param);
451}
452
453static void watchdog_enable(unsigned int cpu)
454{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200455 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400456
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200457 /*
458 * Start the timer first to prevent the NMI watchdog triggering
459 * before the timer has a chance to fire.
460 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100461 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
462 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800463 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400464 HRTIMER_MODE_REL_PINNED);
465
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200466 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000467 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200468 /* Enable the perf event */
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200469 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
470 watchdog_nmi_enable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200471
472 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400473}
474
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000475static void watchdog_disable(unsigned int cpu)
476{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200477 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000478
479 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200480 /*
481 * Disable the perf event first. That prevents that a large delay
482 * between disabling the timer and disabling the perf event causes
483 * the perf NMI to detect a false positive.
484 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000485 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200486 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000487}
488
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200489static void watchdog_cleanup(unsigned int cpu, bool online)
490{
491 watchdog_disable(cpu);
492}
493
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000494static int watchdog_should_run(unsigned int cpu)
495{
496 return __this_cpu_read(hrtimer_interrupts) !=
497 __this_cpu_read(soft_lockup_hrtimer_cnt);
498}
499
500/*
501 * The watchdog thread function - touches the timestamp.
502 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800503 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000504 * default) to reset the softlockup timestamp. If this gets delayed
505 * for more than 2*watchdog_thresh seconds then the debug-printout
506 * triggers in watchdog_timer_fn().
507 */
508static void watchdog(unsigned int cpu)
509{
510 __this_cpu_write(soft_lockup_hrtimer_cnt,
511 __this_cpu_read(hrtimer_interrupts));
512 __touch_watchdog();
513}
Don Zickus58687ac2010-05-07 17:11:44 -0400514
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200515static struct smp_hotplug_thread watchdog_threads = {
516 .store = &softlockup_watchdog,
517 .thread_should_run = watchdog_should_run,
518 .thread_fn = watchdog,
519 .thread_comm = "watchdog/%u",
520 .setup = watchdog_enable,
521 .cleanup = watchdog_cleanup,
522 .park = watchdog_disable,
523 .unpark = watchdog_enable,
524};
525
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200526static void softlockup_update_smpboot_threads(void)
527{
528 lockdep_assert_held(&watchdog_mutex);
529
530 if (!softlockup_threads_initialized)
531 return;
532
533 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
534 &watchdog_allowed_mask);
535 __lockup_detector_cleanup();
536}
537
538/* Temporarily park all watchdog threads */
539static void softlockup_park_all_threads(void)
540{
541 cpumask_clear(&watchdog_allowed_mask);
542 softlockup_update_smpboot_threads();
543}
544
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200545/* Unpark enabled threads */
546static void softlockup_unpark_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200547{
548 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
549 softlockup_update_smpboot_threads();
550}
551
Thomas Gleixner09154982017-09-12 21:37:17 +0200552static void softlockup_reconfigure_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200553{
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200554 watchdog_nmi_reconfigure(false);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200555 softlockup_park_all_threads();
556 set_sample_period();
Thomas Gleixner09154982017-09-12 21:37:17 +0200557 lockup_detector_update_enable();
558 if (watchdog_enabled && watchdog_thresh)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200559 softlockup_unpark_threads();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200560 watchdog_nmi_reconfigure(true);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200561}
562
563/*
564 * Create the watchdog thread infrastructure.
565 *
566 * The threads are not unparked as watchdog_allowed_mask is empty. When
567 * the threads are sucessfully initialized, take the proper locks and
568 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
569 */
570static __init void softlockup_init_threads(void)
571{
572 int ret;
573
574 /*
575 * If sysctl is off and watchdog got disabled on the command line,
576 * nothing to do here.
577 */
Thomas Gleixner09154982017-09-12 21:37:17 +0200578 lockup_detector_update_enable();
579
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200580 if (!IS_ENABLED(CONFIG_SYSCTL) &&
581 !(watchdog_enabled && watchdog_thresh))
582 return;
583
584 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
585 &watchdog_allowed_mask);
586 if (ret) {
587 pr_err("Failed to initialize soft lockup detector threads\n");
588 return;
589 }
590
591 mutex_lock(&watchdog_mutex);
592 softlockup_threads_initialized = true;
Thomas Gleixner09154982017-09-12 21:37:17 +0200593 softlockup_reconfigure_threads();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200594 mutex_unlock(&watchdog_mutex);
595}
596
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200597#else /* CONFIG_SOFTLOCKUP_DETECTOR */
598static inline int watchdog_park_threads(void) { return 0; }
599static inline void watchdog_unpark_threads(void) { }
600static inline int watchdog_enable_all_cpus(void) { return 0; }
601static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200602static inline void softlockup_init_threads(void) { }
Thomas Gleixner09154982017-09-12 21:37:17 +0200603static void softlockup_reconfigure_threads(void)
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200604{
605 watchdog_nmi_reconfigure(false);
Thomas Gleixner09154982017-09-12 21:37:17 +0200606 lockup_detector_update_enable();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200607 watchdog_nmi_reconfigure(true);
608}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200609#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700610
Thomas Gleixner941154b2017-09-12 21:37:04 +0200611static void __lockup_detector_cleanup(void)
612{
613 lockdep_assert_held(&watchdog_mutex);
614 hardlockup_detector_perf_cleanup();
615}
616
617/**
618 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
619 *
620 * Caller must not hold the cpu hotplug rwsem.
621 */
622void lockup_detector_cleanup(void)
623{
624 mutex_lock(&watchdog_mutex);
625 __lockup_detector_cleanup();
626 mutex_unlock(&watchdog_mutex);
627}
628
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200629/**
630 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
631 *
632 * Special interface for parisc. It prevents lockup detector warnings from
633 * the default pm_poweroff() function which busy loops forever.
634 */
635void lockup_detector_soft_poweroff(void)
636{
637 watchdog_enabled = 0;
638}
639
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800640#ifdef CONFIG_SYSCTL
641
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200642/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200643static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400644{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200645 /* Remove impossible cpus to keep sysctl output clean. */
646 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixner09154982017-09-12 21:37:17 +0200647 softlockup_reconfigure_threads();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700648}
649
650/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700651 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
652 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200653 * caller | table->data points to | 'which'
654 * -------------------|----------------------------|--------------------------
655 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
656 * | | SOFT_WATCHDOG_ENABLED
657 * -------------------|----------------------------|--------------------------
658 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
659 * -------------------|----------------------------|--------------------------
660 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700661 */
662static int proc_watchdog_common(int which, struct ctl_table *table, int write,
663 void __user *buffer, size_t *lenp, loff_t *ppos)
664{
Thomas Gleixner09154982017-09-12 21:37:17 +0200665 int err, old, *param = table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400666
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200667 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200668 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000669
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700670 if (!write) {
Thomas Gleixner09154982017-09-12 21:37:17 +0200671 /*
672 * On read synchronize the userspace interface. This is a
673 * racy snapshot.
674 */
675 *param = (watchdog_enabled & which) != 0;
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700676 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
677 } else {
Thomas Gleixner09154982017-09-12 21:37:17 +0200678 old = READ_ONCE(*param);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700679 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Thomas Gleixner09154982017-09-12 21:37:17 +0200680 if (!err && old != READ_ONCE(*param))
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200681 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700682 }
Thomas Gleixner946d1972017-09-12 21:37:01 +0200683 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200684 cpu_hotplug_enable();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700685 return err;
686}
687
688/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700689 * /proc/sys/kernel/watchdog
690 */
691int proc_watchdog(struct ctl_table *table, int write,
692 void __user *buffer, size_t *lenp, loff_t *ppos)
693{
694 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
695 table, write, buffer, lenp, ppos);
696}
697
698/*
699 * /proc/sys/kernel/nmi_watchdog
700 */
701int proc_nmi_watchdog(struct ctl_table *table, int write,
702 void __user *buffer, size_t *lenp, loff_t *ppos)
703{
Thomas Gleixnera994a312017-09-12 21:37:19 +0200704 if (!nmi_watchdog_available && write)
705 return -ENOTSUPP;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700706 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
707 table, write, buffer, lenp, ppos);
708}
709
710/*
711 * /proc/sys/kernel/soft_watchdog
712 */
713int proc_soft_watchdog(struct ctl_table *table, int write,
714 void __user *buffer, size_t *lenp, loff_t *ppos)
715{
716 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
717 table, write, buffer, lenp, ppos);
718}
719
720/*
721 * /proc/sys/kernel/watchdog_thresh
722 */
723int proc_watchdog_thresh(struct ctl_table *table, int write,
724 void __user *buffer, size_t *lenp, loff_t *ppos)
725{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200726 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700727
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200728 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200729 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700730
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200731 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200732 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700733
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200734 if (!err && write && old != READ_ONCE(watchdog_thresh))
735 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700736
Thomas Gleixner946d1972017-09-12 21:37:01 +0200737 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200738 cpu_hotplug_enable();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200739 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400740}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700741
742/*
743 * The cpumask is the mask of possible cpus that the watchdog can run
744 * on, not the mask of cpus it is actually running on. This allows the
745 * user to specify a mask that will include cpus that have not yet
746 * been brought online, if desired.
747 */
748int proc_watchdog_cpumask(struct ctl_table *table, int write,
749 void __user *buffer, size_t *lenp, loff_t *ppos)
750{
751 int err;
752
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200753 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200754 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700755
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700756 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200757 if (!err && write)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200758 proc_watchdog_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200759
Thomas Gleixner946d1972017-09-12 21:37:01 +0200760 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200761 cpu_hotplug_enable();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700762 return err;
763}
Don Zickus58687ac2010-05-07 17:11:44 -0400764#endif /* CONFIG_SYSCTL */
765
Peter Zijlstra004417a2010-11-25 18:38:29 +0100766void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400767{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700768#ifdef CONFIG_NO_HZ_FULL
769 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700770 pr_info("Disabling watchdog on nohz_full cores by default\n");
771 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700772 } else
773 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
774#else
775 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
776#endif
777
Thomas Gleixnera994a312017-09-12 21:37:19 +0200778 if (!watchdog_nmi_probe())
779 nmi_watchdog_available = true;
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200780 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400781}