blob: fd8a998eb1976d7d3e5c19299d15e1b278152ec7 [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{
110 return 0;
111}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200112
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700113void __weak watchdog_nmi_disable(unsigned int cpu)
114{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200115 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700116}
117
Thomas Gleixnera994a312017-09-12 21:37:19 +0200118/* Return 0, if a NMI watchdog is available. Error code otherwise */
119int __weak __init watchdog_nmi_probe(void)
120{
121 return hardlockup_detector_perf_init();
122}
123
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200124/**
125 * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs
126 * @run: If false stop the watchdogs on all enabled CPUs
127 * If true start the watchdogs on all enabled CPUs
128 *
129 * The core call order is:
130 * watchdog_nmi_reconfigure(false);
131 * update_variables();
132 * watchdog_nmi_reconfigure(true);
133 *
134 * The second call which starts the watchdogs again guarantees that the
135 * following variables are stable across the call.
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200136 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700137 * - watchdog_thresh
138 * - watchdog_cpumask
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200139 *
140 * After the call the variables can be changed again.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700141 */
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200142void __weak watchdog_nmi_reconfigure(bool run) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700143
Thomas Gleixner09154982017-09-12 21:37:17 +0200144/**
145 * lockup_detector_update_enable - Update the sysctl enable bit
146 *
147 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
148 * can't race with watchdog_nmi_disable().
149 */
150static void lockup_detector_update_enable(void)
151{
152 watchdog_enabled = 0;
153 if (!watchdog_user_enabled)
154 return;
Thomas Gleixnera994a312017-09-12 21:37:19 +0200155 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
Thomas Gleixner09154982017-09-12 21:37:17 +0200156 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
157 if (soft_watchdog_user_enabled)
158 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
159}
160
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700161#ifdef CONFIG_SOFTLOCKUP_DETECTOR
162
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200163/* Global variables, exported for sysctl */
164unsigned int __read_mostly softlockup_panic =
165 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200166
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800167static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400168
169static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
170static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
171static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
172static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400173static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000174static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
175static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700176static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400177static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700178static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400179
Don Zickus58687ac2010-05-07 17:11:44 -0400180static int __init softlockup_panic_setup(char *str)
181{
182 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400183 return 1;
184}
185__setup("softlockup_panic=", softlockup_panic_setup);
186
187static int __init nowatchdog_setup(char *str)
188{
Thomas Gleixner09154982017-09-12 21:37:17 +0200189 watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400190 return 1;
191}
192__setup("nowatchdog", nowatchdog_setup);
193
Don Zickus58687ac2010-05-07 17:11:44 -0400194static int __init nosoftlockup_setup(char *str)
195{
Thomas Gleixner09154982017-09-12 21:37:17 +0200196 soft_watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400197 return 1;
198}
199__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700200
Aaron Tomlined235872014-06-23 13:22:05 -0700201#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200202int __read_mostly sysctl_softlockup_all_cpu_backtrace;
203
Aaron Tomlined235872014-06-23 13:22:05 -0700204static int __init softlockup_all_cpu_backtrace_setup(char *str)
205{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200206 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700207 return 1;
208}
209__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700210#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400211
Thomas Gleixner941154b2017-09-12 21:37:04 +0200212static void __lockup_detector_cleanup(void);
213
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700214/*
215 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
216 * lockups can have false positives under extreme conditions. So we generally
217 * want a higher threshold for soft lockups than for hard lockups. So we couple
218 * the thresholds with a factor: we make the soft threshold twice the amount of
219 * time the hard threshold is.
220 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200221static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700222{
223 return watchdog_thresh * 2;
224}
Don Zickus58687ac2010-05-07 17:11:44 -0400225
226/*
227 * Returns seconds, approximately. We don't need nanosecond
228 * resolution, and we don't need to waste time with a big divide when
229 * 2^30ns == 1.074s.
230 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900231static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400232{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800233 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400234}
235
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800236static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400237{
238 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700239 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500240 * the divide by 5 is to give hrtimer several chances (two
241 * or three with the current relation between the soft
242 * and hard thresholds) to increment before the
243 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400244 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800245 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200246 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400247}
248
249/* Commands for resetting the watchdog */
250static void __touch_watchdog(void)
251{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900252 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400253}
254
Tejun Heo03e0d462015-12-08 11:28:04 -0500255/**
256 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
257 *
258 * Call when the scheduler may have stalled for legitimate reasons
259 * preventing the watchdog task from executing - e.g. the scheduler
260 * entering idle state. This should only be used for scheduler events.
261 * Use touch_softlockup_watchdog() for everything else.
262 */
263void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400264{
Andrew Morton78611442014-04-18 15:07:12 -0700265 /*
266 * Preemption can be enabled. It doesn't matter which CPU's timestamp
267 * gets zeroed here, so use the raw_ operation.
268 */
269 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400270}
Tejun Heo03e0d462015-12-08 11:28:04 -0500271
272void touch_softlockup_watchdog(void)
273{
274 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500275 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500276}
Ingo Molnar0167c782010-05-13 08:53:33 +0200277EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400278
Don Zickus332fbdb2010-05-07 17:11:45 -0400279void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400280{
281 int cpu;
282
283 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200284 * watchdog_mutex cannpt be taken here, as this might be called
285 * from (soft)interrupt context, so the access to
286 * watchdog_allowed_cpumask might race with a concurrent update.
287 *
288 * The watchdog time stamp can race against a concurrent real
289 * update as well, the only side effect might be a cycle delay for
290 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400291 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200292 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400293 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500294 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400295}
296
Don Zickus58687ac2010-05-07 17:11:44 -0400297void touch_softlockup_watchdog_sync(void)
298{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500299 __this_cpu_write(softlockup_touch_sync, true);
300 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400301}
302
Don Zickus26e09c62010-05-17 18:06:04 -0400303static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400304{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900305 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400306
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800307 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700308 /* Warn about unreasonable delays. */
309 if (time_after(now, touch_ts + get_softlockup_thresh()))
310 return now - touch_ts;
311 }
Don Zickus58687ac2010-05-07 17:11:44 -0400312 return 0;
313}
314
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700315/* watchdog detector functions */
316bool is_hardlockup(void)
317{
318 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
319
320 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
321 return true;
322
323 __this_cpu_write(hrtimer_interrupts_saved, hrint);
324 return false;
325}
326
Don Zickus58687ac2010-05-07 17:11:44 -0400327static void watchdog_interrupt_count(void)
328{
Christoph Lameter909ea962010-12-08 16:22:55 +0100329 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400330}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000331
Don Zickus58687ac2010-05-07 17:11:44 -0400332/* watchdog kicker functions */
333static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
334{
Christoph Lameter909ea962010-12-08 16:22:55 +0100335 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400336 struct pt_regs *regs = get_irq_regs();
337 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700338 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400339
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200340 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800341 return HRTIMER_NORESTART;
342
Don Zickus58687ac2010-05-07 17:11:44 -0400343 /* kick the hardlockup detector */
344 watchdog_interrupt_count();
345
346 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100347 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400348
349 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800350 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400351
352 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100353 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400354 /*
355 * If the time stamp was touched atomically
356 * make sure the scheduler tick is up to date.
357 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100358 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400359 sched_clock_tick();
360 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500361
362 /* Clear the guest paused flag on watchdog reset */
363 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400364 __touch_watchdog();
365 return HRTIMER_RESTART;
366 }
367
368 /* check for a softlockup
369 * This is done by making sure a high priority task is
370 * being scheduled. The task touches the watchdog to
371 * indicate it is getting cpu time. If it hasn't then
372 * this is a good indication some task is hogging the cpu
373 */
Don Zickus26e09c62010-05-17 18:06:04 -0400374 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400375 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500376 /*
377 * If a virtual machine is stopped by the host it can look to
378 * the watchdog like a soft lockup, check to see if the host
379 * stopped the vm before we issue the warning
380 */
381 if (kvm_check_and_clear_guest_paused())
382 return HRTIMER_RESTART;
383
Don Zickus58687ac2010-05-07 17:11:44 -0400384 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700385 if (__this_cpu_read(soft_watchdog_warn) == true) {
386 /*
387 * When multiple processes are causing softlockups the
388 * softlockup detector only warns on the first one
389 * because the code relies on a full quiet cycle to
390 * re-arm. The second process prevents the quiet cycle
391 * and never gets reported. Use task pointers to detect
392 * this.
393 */
394 if (__this_cpu_read(softlockup_task_ptr_saved) !=
395 current) {
396 __this_cpu_write(soft_watchdog_warn, false);
397 __touch_watchdog();
398 }
Don Zickus58687ac2010-05-07 17:11:44 -0400399 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700400 }
Don Zickus58687ac2010-05-07 17:11:44 -0400401
Aaron Tomlined235872014-06-23 13:22:05 -0700402 if (softlockup_all_cpu_backtrace) {
403 /* Prevent multiple soft-lockup reports if one cpu is already
404 * engaged in dumping cpu back traces
405 */
406 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
407 /* Someone else will report us. Let's give up */
408 __this_cpu_write(soft_watchdog_warn, true);
409 return HRTIMER_RESTART;
410 }
411 }
412
Fabian Frederick656c3b72014-08-06 16:04:03 -0700413 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400414 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400415 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700416 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400417 print_modules();
418 print_irqtrace_events(current);
419 if (regs)
420 show_regs(regs);
421 else
422 dump_stack();
423
Aaron Tomlined235872014-06-23 13:22:05 -0700424 if (softlockup_all_cpu_backtrace) {
425 /* Avoid generating two back traces for current
426 * given that one is already made above
427 */
428 trigger_allbutself_cpu_backtrace();
429
430 clear_bit(0, &soft_lockup_nmi_warn);
431 /* Barrier to sync with other cpus */
432 smp_mb__after_atomic();
433 }
434
Josh Hunt69361ee2014-08-08 14:22:31 -0700435 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400436 if (softlockup_panic)
437 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100438 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400439 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100440 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400441
442 return HRTIMER_RESTART;
443}
444
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000445static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400446{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000447 struct sched_param param = { .sched_priority = prio };
448
449 sched_setscheduler(current, policy, &param);
450}
451
452static void watchdog_enable(unsigned int cpu)
453{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200454 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400455
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200456 /*
457 * Start the timer first to prevent the NMI watchdog triggering
458 * before the timer has a chance to fire.
459 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100460 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
461 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800462 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400463 HRTIMER_MODE_REL_PINNED);
464
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200465 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000466 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200467 /* Enable the perf event */
468 watchdog_nmi_enable(cpu);
469
470 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400471}
472
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000473static void watchdog_disable(unsigned int cpu)
474{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200475 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000476
477 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200478 /*
479 * Disable the perf event first. That prevents that a large delay
480 * between disabling the timer and disabling the perf event causes
481 * the perf NMI to detect a false positive.
482 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000483 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200484 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000485}
486
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200487static void watchdog_cleanup(unsigned int cpu, bool online)
488{
489 watchdog_disable(cpu);
490}
491
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000492static int watchdog_should_run(unsigned int cpu)
493{
494 return __this_cpu_read(hrtimer_interrupts) !=
495 __this_cpu_read(soft_lockup_hrtimer_cnt);
496}
497
498/*
499 * The watchdog thread function - touches the timestamp.
500 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800501 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000502 * default) to reset the softlockup timestamp. If this gets delayed
503 * for more than 2*watchdog_thresh seconds then the debug-printout
504 * triggers in watchdog_timer_fn().
505 */
506static void watchdog(unsigned int cpu)
507{
508 __this_cpu_write(soft_lockup_hrtimer_cnt,
509 __this_cpu_read(hrtimer_interrupts));
510 __touch_watchdog();
511}
Don Zickus58687ac2010-05-07 17:11:44 -0400512
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200513static struct smp_hotplug_thread watchdog_threads = {
514 .store = &softlockup_watchdog,
515 .thread_should_run = watchdog_should_run,
516 .thread_fn = watchdog,
517 .thread_comm = "watchdog/%u",
518 .setup = watchdog_enable,
519 .cleanup = watchdog_cleanup,
520 .park = watchdog_disable,
521 .unpark = watchdog_enable,
522};
523
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200524static void softlockup_update_smpboot_threads(void)
525{
526 lockdep_assert_held(&watchdog_mutex);
527
528 if (!softlockup_threads_initialized)
529 return;
530
531 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
532 &watchdog_allowed_mask);
533 __lockup_detector_cleanup();
534}
535
536/* Temporarily park all watchdog threads */
537static void softlockup_park_all_threads(void)
538{
539 cpumask_clear(&watchdog_allowed_mask);
540 softlockup_update_smpboot_threads();
541}
542
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200543/* Unpark enabled threads */
544static void softlockup_unpark_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200545{
546 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
547 softlockup_update_smpboot_threads();
548}
549
Thomas Gleixner09154982017-09-12 21:37:17 +0200550static void softlockup_reconfigure_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200551{
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200552 watchdog_nmi_reconfigure(false);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200553 softlockup_park_all_threads();
554 set_sample_period();
Thomas Gleixner09154982017-09-12 21:37:17 +0200555 lockup_detector_update_enable();
556 if (watchdog_enabled && watchdog_thresh)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200557 softlockup_unpark_threads();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200558 watchdog_nmi_reconfigure(true);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200559}
560
561/*
562 * Create the watchdog thread infrastructure.
563 *
564 * The threads are not unparked as watchdog_allowed_mask is empty. When
565 * the threads are sucessfully initialized, take the proper locks and
566 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
567 */
568static __init void softlockup_init_threads(void)
569{
570 int ret;
571
572 /*
573 * If sysctl is off and watchdog got disabled on the command line,
574 * nothing to do here.
575 */
Thomas Gleixner09154982017-09-12 21:37:17 +0200576 lockup_detector_update_enable();
577
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200578 if (!IS_ENABLED(CONFIG_SYSCTL) &&
579 !(watchdog_enabled && watchdog_thresh))
580 return;
581
582 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
583 &watchdog_allowed_mask);
584 if (ret) {
585 pr_err("Failed to initialize soft lockup detector threads\n");
586 return;
587 }
588
589 mutex_lock(&watchdog_mutex);
590 softlockup_threads_initialized = true;
Thomas Gleixner09154982017-09-12 21:37:17 +0200591 softlockup_reconfigure_threads();
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 */
596static inline int watchdog_park_threads(void) { return 0; }
597static inline void watchdog_unpark_threads(void) { }
598static inline int watchdog_enable_all_cpus(void) { return 0; }
599static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200600static inline void softlockup_init_threads(void) { }
Thomas Gleixner09154982017-09-12 21:37:17 +0200601static void softlockup_reconfigure_threads(void)
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200602{
603 watchdog_nmi_reconfigure(false);
Thomas Gleixner09154982017-09-12 21:37:17 +0200604 lockup_detector_update_enable();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200605 watchdog_nmi_reconfigure(true);
606}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200607#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700608
Thomas Gleixner941154b2017-09-12 21:37:04 +0200609static void __lockup_detector_cleanup(void)
610{
611 lockdep_assert_held(&watchdog_mutex);
612 hardlockup_detector_perf_cleanup();
613}
614
615/**
616 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
617 *
618 * Caller must not hold the cpu hotplug rwsem.
619 */
620void lockup_detector_cleanup(void)
621{
622 mutex_lock(&watchdog_mutex);
623 __lockup_detector_cleanup();
624 mutex_unlock(&watchdog_mutex);
625}
626
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200627/**
628 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
629 *
630 * Special interface for parisc. It prevents lockup detector warnings from
631 * the default pm_poweroff() function which busy loops forever.
632 */
633void lockup_detector_soft_poweroff(void)
634{
635 watchdog_enabled = 0;
636}
637
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800638#ifdef CONFIG_SYSCTL
639
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200640/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200641static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400642{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200643 /* Remove impossible cpus to keep sysctl output clean. */
644 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixner09154982017-09-12 21:37:17 +0200645 softlockup_reconfigure_threads();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700646}
647
648/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700649 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
650 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200651 * caller | table->data points to | 'which'
652 * -------------------|----------------------------|--------------------------
653 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
654 * | | SOFT_WATCHDOG_ENABLED
655 * -------------------|----------------------------|--------------------------
656 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
657 * -------------------|----------------------------|--------------------------
658 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700659 */
660static int proc_watchdog_common(int which, struct ctl_table *table, int write,
661 void __user *buffer, size_t *lenp, loff_t *ppos)
662{
Thomas Gleixner09154982017-09-12 21:37:17 +0200663 int err, old, *param = table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400664
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200665 cpu_hotplug_disable();
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);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200682 cpu_hotplug_enable();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700683 return err;
684}
685
686/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700687 * /proc/sys/kernel/watchdog
688 */
689int proc_watchdog(struct ctl_table *table, int write,
690 void __user *buffer, size_t *lenp, loff_t *ppos)
691{
692 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
693 table, write, buffer, lenp, ppos);
694}
695
696/*
697 * /proc/sys/kernel/nmi_watchdog
698 */
699int proc_nmi_watchdog(struct ctl_table *table, int write,
700 void __user *buffer, size_t *lenp, loff_t *ppos)
701{
Thomas Gleixnera994a312017-09-12 21:37:19 +0200702 if (!nmi_watchdog_available && write)
703 return -ENOTSUPP;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700704 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
705 table, write, buffer, lenp, ppos);
706}
707
708/*
709 * /proc/sys/kernel/soft_watchdog
710 */
711int proc_soft_watchdog(struct ctl_table *table, int write,
712 void __user *buffer, size_t *lenp, loff_t *ppos)
713{
714 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
715 table, write, buffer, lenp, ppos);
716}
717
718/*
719 * /proc/sys/kernel/watchdog_thresh
720 */
721int proc_watchdog_thresh(struct ctl_table *table, int write,
722 void __user *buffer, size_t *lenp, loff_t *ppos)
723{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200724 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700725
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200726 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200727 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700728
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200729 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200730 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700731
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200732 if (!err && write && old != READ_ONCE(watchdog_thresh))
733 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700734
Thomas Gleixner946d1972017-09-12 21:37:01 +0200735 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200736 cpu_hotplug_enable();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200737 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400738}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700739
740/*
741 * The cpumask is the mask of possible cpus that the watchdog can run
742 * on, not the mask of cpus it is actually running on. This allows the
743 * user to specify a mask that will include cpus that have not yet
744 * been brought online, if desired.
745 */
746int proc_watchdog_cpumask(struct ctl_table *table, int write,
747 void __user *buffer, size_t *lenp, loff_t *ppos)
748{
749 int err;
750
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200751 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200752 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700753
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700754 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200755 if (!err && write)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200756 proc_watchdog_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200757
Thomas Gleixner946d1972017-09-12 21:37:01 +0200758 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200759 cpu_hotplug_enable();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700760 return err;
761}
Don Zickus58687ac2010-05-07 17:11:44 -0400762#endif /* CONFIG_SYSCTL */
763
Peter Zijlstra004417a2010-11-25 18:38:29 +0100764void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400765{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700766#ifdef CONFIG_NO_HZ_FULL
767 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700768 pr_info("Disabling watchdog on nohz_full cores by default\n");
769 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700770 } else
771 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
772#else
773 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
774#endif
775
Thomas Gleixnera994a312017-09-12 21:37:19 +0200776 if (!watchdog_nmi_probe())
777 nmi_watchdog_available = true;
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200778 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400779}