blob: b81f777838d5ab30f20c6a46228803a71251dc27 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Don Zickus58687ac2010-05-07 17:11:44 -04002/*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -05007 * Note: Most of this code is borrowed heavily from the original softlockup
8 * detector, so thanks to Ingo for the initial implementation.
9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
Don Zickus58687ac2010-05-07 17:11:44 -040010 * to those contributors as well.
11 */
12
Kefeng Wang5f92a7b2017-07-14 14:49:46 -070013#define pr_fmt(fmt) "watchdog: " fmt
Andrew Morton45019802012-03-23 15:01:55 -070014
Don Zickus58687ac2010-05-07 17:11:44 -040015#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
Don Zickus58687ac2010-05-07 17:11:44 -040019#include <linux/module.h>
20#include <linux/sysctl.h>
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070021#include <linux/tick.h>
Ingo Molnare6017572017-02-01 16:36:40 +010022#include <linux/sched/clock.h>
Ingo Molnarb17b0152017-02-08 18:51:35 +010023#include <linux/sched/debug.h>
Frederic Weisbecker78634062017-10-27 04:42:28 +020024#include <linux/sched/isolation.h>
Peter Zijlstra9cf57732018-06-07 10:52:03 +020025#include <linux/stop_machine.h>
Don Zickus58687ac2010-05-07 17:11:44 -040026
27#include <asm/irq_regs.h>
Eric B Munson5d1c0f42012-03-10 14:37:28 -050028#include <linux/kvm_para.h>
Don Zickus58687ac2010-05-07 17:11:44 -040029
Thomas Gleixner946d1972017-09-12 21:37:01 +020030static DEFINE_MUTEX(watchdog_mutex);
Peter Zijlstraab992dc2015-05-18 11:31:50 +020031
Nicholas Piggin05a4a952017-07-12 14:35:46 -070032#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
Thomas Gleixner09154982017-09-12 21:37:17 +020033# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
34# define NMI_WATCHDOG_DEFAULT 1
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070035#else
Thomas Gleixner09154982017-09-12 21:37:17 +020036# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37# define NMI_WATCHDOG_DEFAULT 0
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070038#endif
Nicholas Piggin05a4a952017-07-12 14:35:46 -070039
Thomas Gleixner09154982017-09-12 21:37:17 +020040unsigned long __read_mostly watchdog_enabled;
41int __read_mostly watchdog_user_enabled = 1;
42int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43int __read_mostly soft_watchdog_user_enabled = 1;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020044int __read_mostly watchdog_thresh = 10;
Thomas Gleixnera994a312017-09-12 21:37:19 +020045int __read_mostly nmi_watchdog_available;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020046
47struct cpumask watchdog_allowed_mask __read_mostly;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020048
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{
Thomas Gleixner09154982017-09-12 21:37:17 +020068 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070069}
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))
Thomas Gleixner09154982017-09-12 21:37:17 +020078 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070079 else if (!strncmp(str, "1", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020080 nmi_watchdog_user_enabled = 1;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070081 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{
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200107 hardlockup_detector_perf_enable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700108 return 0;
109}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200110
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700111void __weak watchdog_nmi_disable(unsigned int cpu)
112{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200113 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700114}
115
Thomas Gleixnera994a312017-09-12 21:37:19 +0200116/* Return 0, if a NMI watchdog is available. Error code otherwise */
117int __weak __init watchdog_nmi_probe(void)
118{
119 return hardlockup_detector_perf_init();
120}
121
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200122/**
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200123 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200124 *
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200125 * The reconfiguration steps are:
126 * watchdog_nmi_stop();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200127 * update_variables();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200128 * watchdog_nmi_start();
129 */
130void __weak watchdog_nmi_stop(void) { }
131
132/**
133 * watchdog_nmi_start - Start the watchdog after reconfiguration
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200134 *
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200135 * Counterpart to watchdog_nmi_stop().
136 *
137 * The following variables have been updated in update_variables() and
138 * contain the currently valid configuration:
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200139 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700140 * - watchdog_thresh
141 * - watchdog_cpumask
Nicholas Piggina10a8422017-07-12 14:35:49 -0700142 */
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200143void __weak watchdog_nmi_start(void) { }
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
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200168static bool softlockup_initialized __read_mostly;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800169static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400170
171static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400172static 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
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200333/*
334 * The watchdog thread function - touches the timestamp.
335 *
336 * It only runs once every sample_period seconds (4 seconds by
337 * default) to reset the softlockup timestamp. If this gets delayed
338 * for more than 2*watchdog_thresh seconds then the debug-printout
339 * triggers in watchdog_timer_fn().
340 */
341static int softlockup_fn(void *data)
342{
343 __this_cpu_write(soft_lockup_hrtimer_cnt,
344 __this_cpu_read(hrtimer_interrupts));
345 __touch_watchdog();
346
347 return 0;
348}
349
350static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
351
Don Zickus58687ac2010-05-07 17:11:44 -0400352/* watchdog kicker functions */
353static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
354{
Christoph Lameter909ea962010-12-08 16:22:55 +0100355 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400356 struct pt_regs *regs = get_irq_regs();
357 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700358 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400359
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200360 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800361 return HRTIMER_NORESTART;
362
Don Zickus58687ac2010-05-07 17:11:44 -0400363 /* kick the hardlockup detector */
364 watchdog_interrupt_count();
365
366 /* kick the softlockup detector */
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200367 stop_one_cpu_nowait(smp_processor_id(),
368 softlockup_fn, NULL,
369 this_cpu_ptr(&softlockup_stop_work));
Don Zickus58687ac2010-05-07 17:11:44 -0400370
371 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800372 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400373
374 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100375 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400376 /*
377 * If the time stamp was touched atomically
378 * make sure the scheduler tick is up to date.
379 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100380 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400381 sched_clock_tick();
382 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500383
384 /* Clear the guest paused flag on watchdog reset */
385 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400386 __touch_watchdog();
387 return HRTIMER_RESTART;
388 }
389
390 /* check for a softlockup
391 * This is done by making sure a high priority task is
392 * being scheduled. The task touches the watchdog to
393 * indicate it is getting cpu time. If it hasn't then
394 * this is a good indication some task is hogging the cpu
395 */
Don Zickus26e09c62010-05-17 18:06:04 -0400396 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400397 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500398 /*
399 * If a virtual machine is stopped by the host it can look to
400 * the watchdog like a soft lockup, check to see if the host
401 * stopped the vm before we issue the warning
402 */
403 if (kvm_check_and_clear_guest_paused())
404 return HRTIMER_RESTART;
405
Don Zickus58687ac2010-05-07 17:11:44 -0400406 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700407 if (__this_cpu_read(soft_watchdog_warn) == true) {
408 /*
409 * When multiple processes are causing softlockups the
410 * softlockup detector only warns on the first one
411 * because the code relies on a full quiet cycle to
412 * re-arm. The second process prevents the quiet cycle
413 * and never gets reported. Use task pointers to detect
414 * this.
415 */
416 if (__this_cpu_read(softlockup_task_ptr_saved) !=
417 current) {
418 __this_cpu_write(soft_watchdog_warn, false);
419 __touch_watchdog();
420 }
Don Zickus58687ac2010-05-07 17:11:44 -0400421 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700422 }
Don Zickus58687ac2010-05-07 17:11:44 -0400423
Aaron Tomlined235872014-06-23 13:22:05 -0700424 if (softlockup_all_cpu_backtrace) {
425 /* Prevent multiple soft-lockup reports if one cpu is already
426 * engaged in dumping cpu back traces
427 */
428 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
429 /* Someone else will report us. Let's give up */
430 __this_cpu_write(soft_watchdog_warn, true);
431 return HRTIMER_RESTART;
432 }
433 }
434
Fabian Frederick656c3b72014-08-06 16:04:03 -0700435 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400436 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400437 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700438 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400439 print_modules();
440 print_irqtrace_events(current);
441 if (regs)
442 show_regs(regs);
443 else
444 dump_stack();
445
Aaron Tomlined235872014-06-23 13:22:05 -0700446 if (softlockup_all_cpu_backtrace) {
447 /* Avoid generating two back traces for current
448 * given that one is already made above
449 */
450 trigger_allbutself_cpu_backtrace();
451
452 clear_bit(0, &soft_lockup_nmi_warn);
453 /* Barrier to sync with other cpus */
454 smp_mb__after_atomic();
455 }
456
Josh Hunt69361ee2014-08-08 14:22:31 -0700457 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400458 if (softlockup_panic)
459 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100460 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400461 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100462 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400463
464 return HRTIMER_RESTART;
465}
466
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000467static void watchdog_enable(unsigned int cpu)
468{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200469 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400470
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200471 WARN_ON_ONCE(cpu != smp_processor_id());
472
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200473 /*
474 * Start the timer first to prevent the NMI watchdog triggering
475 * before the timer has a chance to fire.
476 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100477 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
478 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800479 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400480 HRTIMER_MODE_REL_PINNED);
481
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200482 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000483 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200484 /* Enable the perf event */
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200485 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
486 watchdog_nmi_enable(cpu);
Don Zickus58687ac2010-05-07 17:11:44 -0400487}
488
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000489static void watchdog_disable(unsigned int cpu)
490{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200491 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000492
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200493 WARN_ON_ONCE(cpu != smp_processor_id());
494
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200495 /*
496 * Disable the perf event first. That prevents that a large delay
497 * between disabling the timer and disabling the perf event causes
498 * the perf NMI to detect a false positive.
499 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000500 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200501 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000502}
503
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200504static int softlockup_stop_fn(void *data)
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200505{
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200506 watchdog_disable(smp_processor_id());
507 return 0;
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200508}
509
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200510static void softlockup_stop_all(void)
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000511{
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200512 int cpu;
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000513
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200514 if (!softlockup_initialized)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200515 return;
516
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200517 for_each_cpu(cpu, &watchdog_allowed_mask)
518 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200519
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200520 cpumask_clear(&watchdog_allowed_mask);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200521}
522
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200523static int softlockup_start_fn(void *data)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200524{
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200525 watchdog_enable(smp_processor_id());
526 return 0;
527}
528
529static void softlockup_start_all(void)
530{
531 int cpu;
532
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200533 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200534 for_each_cpu(cpu, &watchdog_allowed_mask)
535 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
536}
537
538int lockup_detector_online_cpu(unsigned int cpu)
539{
540 watchdog_enable(cpu);
541 return 0;
542}
543
544int lockup_detector_offline_cpu(unsigned int cpu)
545{
546 watchdog_disable(cpu);
547 return 0;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200548}
549
Thomas Gleixner55871852017-10-04 10:03:04 +0200550static void lockup_detector_reconfigure(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200551{
Thomas Gleixnere31d6882017-10-03 16:37:53 +0200552 cpus_read_lock();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200553 watchdog_nmi_stop();
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200554
555 softlockup_stop_all();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200556 set_sample_period();
Thomas Gleixner09154982017-09-12 21:37:17 +0200557 lockup_detector_update_enable();
558 if (watchdog_enabled && watchdog_thresh)
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200559 softlockup_start_all();
560
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200561 watchdog_nmi_start();
Thomas Gleixnere31d6882017-10-03 16:37:53 +0200562 cpus_read_unlock();
563 /*
564 * Must be called outside the cpus locked section to prevent
565 * recursive locking in the perf code.
566 */
567 __lockup_detector_cleanup();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200568}
569
570/*
Thomas Gleixner55871852017-10-04 10:03:04 +0200571 * Create the watchdog thread infrastructure and configure the detector(s).
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200572 *
573 * The threads are not unparked as watchdog_allowed_mask is empty. When
574 * the threads are sucessfully initialized, take the proper locks and
575 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
576 */
Thomas Gleixner55871852017-10-04 10:03:04 +0200577static __init void lockup_detector_setup(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200578{
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200579 /*
580 * If sysctl is off and watchdog got disabled on the command line,
581 * nothing to do here.
582 */
Thomas Gleixner09154982017-09-12 21:37:17 +0200583 lockup_detector_update_enable();
584
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200585 if (!IS_ENABLED(CONFIG_SYSCTL) &&
586 !(watchdog_enabled && watchdog_thresh))
587 return;
588
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200589 mutex_lock(&watchdog_mutex);
Thomas Gleixner55871852017-10-04 10:03:04 +0200590 lockup_detector_reconfigure();
Peter Zijlstra9cf57732018-06-07 10:52:03 +0200591 softlockup_initialized = true;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200592 mutex_unlock(&watchdog_mutex);
593}
594
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200595#else /* CONFIG_SOFTLOCKUP_DETECTOR */
Thomas Gleixner55871852017-10-04 10:03:04 +0200596static void lockup_detector_reconfigure(void)
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200597{
Thomas Gleixnere31d6882017-10-03 16:37:53 +0200598 cpus_read_lock();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200599 watchdog_nmi_stop();
Thomas Gleixner09154982017-09-12 21:37:17 +0200600 lockup_detector_update_enable();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200601 watchdog_nmi_start();
Thomas Gleixnere31d6882017-10-03 16:37:53 +0200602 cpus_read_unlock();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200603}
Thomas Gleixner55871852017-10-04 10:03:04 +0200604static inline void lockup_detector_setup(void)
Thomas Gleixner34ddaa32017-10-03 16:39:02 +0200605{
Thomas Gleixner55871852017-10-04 10:03:04 +0200606 lockup_detector_reconfigure();
Thomas Gleixner34ddaa32017-10-03 16:39:02 +0200607}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200608#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700609
Thomas Gleixner941154b2017-09-12 21:37:04 +0200610static void __lockup_detector_cleanup(void)
611{
612 lockdep_assert_held(&watchdog_mutex);
613 hardlockup_detector_perf_cleanup();
614}
615
616/**
617 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
618 *
619 * Caller must not hold the cpu hotplug rwsem.
620 */
621void lockup_detector_cleanup(void)
622{
623 mutex_lock(&watchdog_mutex);
624 __lockup_detector_cleanup();
625 mutex_unlock(&watchdog_mutex);
626}
627
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200628/**
629 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
630 *
631 * Special interface for parisc. It prevents lockup detector warnings from
632 * the default pm_poweroff() function which busy loops forever.
633 */
634void lockup_detector_soft_poweroff(void)
635{
636 watchdog_enabled = 0;
637}
638
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800639#ifdef CONFIG_SYSCTL
640
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200641/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200642static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400643{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200644 /* Remove impossible cpus to keep sysctl output clean. */
645 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixner55871852017-10-04 10:03:04 +0200646 lockup_detector_reconfigure();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700647}
648
649/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700650 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
651 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200652 * caller | table->data points to | 'which'
653 * -------------------|----------------------------|--------------------------
654 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
655 * | | SOFT_WATCHDOG_ENABLED
656 * -------------------|----------------------------|--------------------------
657 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
658 * -------------------|----------------------------|--------------------------
659 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700660 */
661static int proc_watchdog_common(int which, struct ctl_table *table, int write,
662 void __user *buffer, size_t *lenp, loff_t *ppos)
663{
Thomas Gleixner09154982017-09-12 21:37:17 +0200664 int err, old, *param = table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400665
Thomas Gleixner946d1972017-09-12 21:37:01 +0200666 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000667
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700668 if (!write) {
Thomas Gleixner09154982017-09-12 21:37:17 +0200669 /*
670 * On read synchronize the userspace interface. This is a
671 * racy snapshot.
672 */
673 *param = (watchdog_enabled & which) != 0;
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700674 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
675 } else {
Thomas Gleixner09154982017-09-12 21:37:17 +0200676 old = READ_ONCE(*param);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700677 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Thomas Gleixner09154982017-09-12 21:37:17 +0200678 if (!err && old != READ_ONCE(*param))
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200679 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700680 }
Thomas Gleixner946d1972017-09-12 21:37:01 +0200681 mutex_unlock(&watchdog_mutex);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700682 return err;
683}
684
685/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700686 * /proc/sys/kernel/watchdog
687 */
688int proc_watchdog(struct ctl_table *table, int write,
689 void __user *buffer, size_t *lenp, loff_t *ppos)
690{
691 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
692 table, write, buffer, lenp, ppos);
693}
694
695/*
696 * /proc/sys/kernel/nmi_watchdog
697 */
698int proc_nmi_watchdog(struct ctl_table *table, int write,
699 void __user *buffer, size_t *lenp, loff_t *ppos)
700{
Thomas Gleixnera994a312017-09-12 21:37:19 +0200701 if (!nmi_watchdog_available && write)
702 return -ENOTSUPP;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700703 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
704 table, write, buffer, lenp, ppos);
705}
706
707/*
708 * /proc/sys/kernel/soft_watchdog
709 */
710int proc_soft_watchdog(struct ctl_table *table, int write,
711 void __user *buffer, size_t *lenp, loff_t *ppos)
712{
713 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
714 table, write, buffer, lenp, ppos);
715}
716
717/*
718 * /proc/sys/kernel/watchdog_thresh
719 */
720int proc_watchdog_thresh(struct ctl_table *table, int write,
721 void __user *buffer, size_t *lenp, loff_t *ppos)
722{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200723 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700724
Thomas Gleixner946d1972017-09-12 21:37:01 +0200725 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700726
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200727 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200728 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700729
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200730 if (!err && write && old != READ_ONCE(watchdog_thresh))
731 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700732
Thomas Gleixner946d1972017-09-12 21:37:01 +0200733 mutex_unlock(&watchdog_mutex);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200734 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400735}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700736
737/*
738 * The cpumask is the mask of possible cpus that the watchdog can run
739 * on, not the mask of cpus it is actually running on. This allows the
740 * user to specify a mask that will include cpus that have not yet
741 * been brought online, if desired.
742 */
743int proc_watchdog_cpumask(struct ctl_table *table, int write,
744 void __user *buffer, size_t *lenp, loff_t *ppos)
745{
746 int err;
747
Thomas Gleixner946d1972017-09-12 21:37:01 +0200748 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700749
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700750 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200751 if (!err && write)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200752 proc_watchdog_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200753
Thomas Gleixner946d1972017-09-12 21:37:01 +0200754 mutex_unlock(&watchdog_mutex);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700755 return err;
756}
Don Zickus58687ac2010-05-07 17:11:44 -0400757#endif /* CONFIG_SYSCTL */
758
Peter Zijlstra004417a2010-11-25 18:38:29 +0100759void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400760{
Frederic Weisbecker13316b32017-10-27 04:42:29 +0200761 if (tick_nohz_full_enabled())
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700762 pr_info("Disabling watchdog on nohz_full cores by default\n");
Frederic Weisbecker13316b32017-10-27 04:42:29 +0200763
Frederic Weisbeckerde201552017-10-27 04:42:35 +0200764 cpumask_copy(&watchdog_cpumask,
765 housekeeping_cpumask(HK_FLAG_TIMER));
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700766
Thomas Gleixnera994a312017-09-12 21:37:19 +0200767 if (!watchdog_nmi_probe())
768 nmi_watchdog_available = true;
Thomas Gleixner55871852017-10-04 10:03:04 +0200769 lockup_detector_setup();
Don Zickus58687ac2010-05-07 17:11:44 -0400770}