blob: 84886319d7b09005f403c3c7bc1911712b780a3a [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;
47
48struct cpumask watchdog_allowed_mask __read_mostly;
49static bool softlockup_threads_initialized __read_mostly;
50
51struct cpumask watchdog_cpumask __read_mostly;
52unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
53
Nicholas Piggin05a4a952017-07-12 14:35:46 -070054#ifdef CONFIG_HARDLOCKUP_DETECTOR
Nicholas Piggin05a4a952017-07-12 14:35:46 -070055/*
56 * Should we panic when a soft-lockup or hard-lockup occurs:
57 */
58unsigned int __read_mostly hardlockup_panic =
59 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
60/*
61 * We may not want to enable hard lockup detection by default in all cases,
62 * for example when running the kernel as a guest on a hypervisor. In these
63 * cases this function can be called to disable hard lockup detection. This
64 * function should only be executed once by the boot processor before the
65 * kernel command line parameters are parsed, because otherwise it is not
66 * possible to override this in hardlockup_panic_setup().
67 */
Thomas Gleixner7a355822017-09-12 21:37:02 +020068void __init hardlockup_detector_disable(void)
Nicholas Piggin05a4a952017-07-12 14:35:46 -070069{
Thomas Gleixner09154982017-09-12 21:37:17 +020070 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070071}
72
73static int __init hardlockup_panic_setup(char *str)
74{
75 if (!strncmp(str, "panic", 5))
76 hardlockup_panic = 1;
77 else if (!strncmp(str, "nopanic", 7))
78 hardlockup_panic = 0;
79 else if (!strncmp(str, "0", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020080 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070081 else if (!strncmp(str, "1", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020082 nmi_watchdog_user_enabled = 1;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070083 return 1;
84}
85__setup("nmi_watchdog=", hardlockup_panic_setup);
86
Thomas Gleixner368a7e22017-09-12 21:37:07 +020087# ifdef CONFIG_SMP
88int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070089
Thomas Gleixner368a7e22017-09-12 21:37:07 +020090static int __init hardlockup_all_cpu_backtrace_setup(char *str)
91{
92 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
93 return 1;
94}
95__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
96# endif /* CONFIG_SMP */
97#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -070098
Ulrich Obergfellec6a9062015-09-04 15:45:28 -070099/*
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700100 * These functions can be overridden if an architecture implements its
101 * own hardlockup detector.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700102 *
103 * watchdog_nmi_enable/disable can be implemented to start and stop when
104 * softlockup watchdog threads start and stop. The arch must select the
105 * SOFTLOCKUP_DETECTOR Kconfig.
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700106 */
107int __weak watchdog_nmi_enable(unsigned int cpu)
108{
109 return 0;
110}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200111
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700112void __weak watchdog_nmi_disable(unsigned int cpu)
113{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200114 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700115}
116
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200117/**
118 * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs
119 * @run: If false stop the watchdogs on all enabled CPUs
120 * If true start the watchdogs on all enabled CPUs
121 *
122 * The core call order is:
123 * watchdog_nmi_reconfigure(false);
124 * update_variables();
125 * watchdog_nmi_reconfigure(true);
126 *
127 * The second call which starts the watchdogs again guarantees that the
128 * following variables are stable across the call.
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200129 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700130 * - watchdog_thresh
131 * - watchdog_cpumask
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200132 *
133 * After the call the variables can be changed again.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700134 */
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200135void __weak watchdog_nmi_reconfigure(bool run) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700136
Thomas Gleixner09154982017-09-12 21:37:17 +0200137/**
138 * lockup_detector_update_enable - Update the sysctl enable bit
139 *
140 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
141 * can't race with watchdog_nmi_disable().
142 */
143static void lockup_detector_update_enable(void)
144{
145 watchdog_enabled = 0;
146 if (!watchdog_user_enabled)
147 return;
148 if (nmi_watchdog_user_enabled)
149 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
150 if (soft_watchdog_user_enabled)
151 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
152}
153
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700154#ifdef CONFIG_SOFTLOCKUP_DETECTOR
155
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200156/* Global variables, exported for sysctl */
157unsigned int __read_mostly softlockup_panic =
158 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200159
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800160static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400161
162static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
163static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
164static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
165static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400166static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000167static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
168static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700169static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400170static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700171static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400172
Don Zickus58687ac2010-05-07 17:11:44 -0400173static int __init softlockup_panic_setup(char *str)
174{
175 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400176 return 1;
177}
178__setup("softlockup_panic=", softlockup_panic_setup);
179
180static int __init nowatchdog_setup(char *str)
181{
Thomas Gleixner09154982017-09-12 21:37:17 +0200182 watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400183 return 1;
184}
185__setup("nowatchdog", nowatchdog_setup);
186
Don Zickus58687ac2010-05-07 17:11:44 -0400187static int __init nosoftlockup_setup(char *str)
188{
Thomas Gleixner09154982017-09-12 21:37:17 +0200189 soft_watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400190 return 1;
191}
192__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700193
Aaron Tomlined235872014-06-23 13:22:05 -0700194#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200195int __read_mostly sysctl_softlockup_all_cpu_backtrace;
196
Aaron Tomlined235872014-06-23 13:22:05 -0700197static int __init softlockup_all_cpu_backtrace_setup(char *str)
198{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200199 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700200 return 1;
201}
202__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700203#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400204
Thomas Gleixner941154b2017-09-12 21:37:04 +0200205static void __lockup_detector_cleanup(void);
206
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700207/*
208 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
209 * lockups can have false positives under extreme conditions. So we generally
210 * want a higher threshold for soft lockups than for hard lockups. So we couple
211 * the thresholds with a factor: we make the soft threshold twice the amount of
212 * time the hard threshold is.
213 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200214static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700215{
216 return watchdog_thresh * 2;
217}
Don Zickus58687ac2010-05-07 17:11:44 -0400218
219/*
220 * Returns seconds, approximately. We don't need nanosecond
221 * resolution, and we don't need to waste time with a big divide when
222 * 2^30ns == 1.074s.
223 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900224static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400225{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800226 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400227}
228
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800229static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400230{
231 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700232 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500233 * the divide by 5 is to give hrtimer several chances (two
234 * or three with the current relation between the soft
235 * and hard thresholds) to increment before the
236 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400237 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800238 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200239 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400240}
241
242/* Commands for resetting the watchdog */
243static void __touch_watchdog(void)
244{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900245 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400246}
247
Tejun Heo03e0d462015-12-08 11:28:04 -0500248/**
249 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
250 *
251 * Call when the scheduler may have stalled for legitimate reasons
252 * preventing the watchdog task from executing - e.g. the scheduler
253 * entering idle state. This should only be used for scheduler events.
254 * Use touch_softlockup_watchdog() for everything else.
255 */
256void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400257{
Andrew Morton78611442014-04-18 15:07:12 -0700258 /*
259 * Preemption can be enabled. It doesn't matter which CPU's timestamp
260 * gets zeroed here, so use the raw_ operation.
261 */
262 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400263}
Tejun Heo03e0d462015-12-08 11:28:04 -0500264
265void touch_softlockup_watchdog(void)
266{
267 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500268 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500269}
Ingo Molnar0167c782010-05-13 08:53:33 +0200270EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400271
Don Zickus332fbdb2010-05-07 17:11:45 -0400272void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400273{
274 int cpu;
275
276 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200277 * watchdog_mutex cannpt be taken here, as this might be called
278 * from (soft)interrupt context, so the access to
279 * watchdog_allowed_cpumask might race with a concurrent update.
280 *
281 * The watchdog time stamp can race against a concurrent real
282 * update as well, the only side effect might be a cycle delay for
283 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400284 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200285 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400286 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500287 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400288}
289
Don Zickus58687ac2010-05-07 17:11:44 -0400290void touch_softlockup_watchdog_sync(void)
291{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500292 __this_cpu_write(softlockup_touch_sync, true);
293 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400294}
295
Don Zickus26e09c62010-05-17 18:06:04 -0400296static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400297{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900298 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400299
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800300 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700301 /* Warn about unreasonable delays. */
302 if (time_after(now, touch_ts + get_softlockup_thresh()))
303 return now - touch_ts;
304 }
Don Zickus58687ac2010-05-07 17:11:44 -0400305 return 0;
306}
307
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700308/* watchdog detector functions */
309bool is_hardlockup(void)
310{
311 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
312
313 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
314 return true;
315
316 __this_cpu_write(hrtimer_interrupts_saved, hrint);
317 return false;
318}
319
Don Zickus58687ac2010-05-07 17:11:44 -0400320static void watchdog_interrupt_count(void)
321{
Christoph Lameter909ea962010-12-08 16:22:55 +0100322 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400323}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000324
Don Zickus58687ac2010-05-07 17:11:44 -0400325/* watchdog kicker functions */
326static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
327{
Christoph Lameter909ea962010-12-08 16:22:55 +0100328 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400329 struct pt_regs *regs = get_irq_regs();
330 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700331 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400332
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200333 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800334 return HRTIMER_NORESTART;
335
Don Zickus58687ac2010-05-07 17:11:44 -0400336 /* kick the hardlockup detector */
337 watchdog_interrupt_count();
338
339 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100340 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400341
342 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800343 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400344
345 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100346 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400347 /*
348 * If the time stamp was touched atomically
349 * make sure the scheduler tick is up to date.
350 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100351 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400352 sched_clock_tick();
353 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500354
355 /* Clear the guest paused flag on watchdog reset */
356 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400357 __touch_watchdog();
358 return HRTIMER_RESTART;
359 }
360
361 /* check for a softlockup
362 * This is done by making sure a high priority task is
363 * being scheduled. The task touches the watchdog to
364 * indicate it is getting cpu time. If it hasn't then
365 * this is a good indication some task is hogging the cpu
366 */
Don Zickus26e09c62010-05-17 18:06:04 -0400367 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400368 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500369 /*
370 * If a virtual machine is stopped by the host it can look to
371 * the watchdog like a soft lockup, check to see if the host
372 * stopped the vm before we issue the warning
373 */
374 if (kvm_check_and_clear_guest_paused())
375 return HRTIMER_RESTART;
376
Don Zickus58687ac2010-05-07 17:11:44 -0400377 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700378 if (__this_cpu_read(soft_watchdog_warn) == true) {
379 /*
380 * When multiple processes are causing softlockups the
381 * softlockup detector only warns on the first one
382 * because the code relies on a full quiet cycle to
383 * re-arm. The second process prevents the quiet cycle
384 * and never gets reported. Use task pointers to detect
385 * this.
386 */
387 if (__this_cpu_read(softlockup_task_ptr_saved) !=
388 current) {
389 __this_cpu_write(soft_watchdog_warn, false);
390 __touch_watchdog();
391 }
Don Zickus58687ac2010-05-07 17:11:44 -0400392 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700393 }
Don Zickus58687ac2010-05-07 17:11:44 -0400394
Aaron Tomlined235872014-06-23 13:22:05 -0700395 if (softlockup_all_cpu_backtrace) {
396 /* Prevent multiple soft-lockup reports if one cpu is already
397 * engaged in dumping cpu back traces
398 */
399 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
400 /* Someone else will report us. Let's give up */
401 __this_cpu_write(soft_watchdog_warn, true);
402 return HRTIMER_RESTART;
403 }
404 }
405
Fabian Frederick656c3b72014-08-06 16:04:03 -0700406 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400407 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400408 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700409 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400410 print_modules();
411 print_irqtrace_events(current);
412 if (regs)
413 show_regs(regs);
414 else
415 dump_stack();
416
Aaron Tomlined235872014-06-23 13:22:05 -0700417 if (softlockup_all_cpu_backtrace) {
418 /* Avoid generating two back traces for current
419 * given that one is already made above
420 */
421 trigger_allbutself_cpu_backtrace();
422
423 clear_bit(0, &soft_lockup_nmi_warn);
424 /* Barrier to sync with other cpus */
425 smp_mb__after_atomic();
426 }
427
Josh Hunt69361ee2014-08-08 14:22:31 -0700428 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400429 if (softlockup_panic)
430 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100431 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400432 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100433 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400434
435 return HRTIMER_RESTART;
436}
437
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000438static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400439{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000440 struct sched_param param = { .sched_priority = prio };
441
442 sched_setscheduler(current, policy, &param);
443}
444
445static void watchdog_enable(unsigned int cpu)
446{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200447 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400448
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200449 /*
450 * Start the timer first to prevent the NMI watchdog triggering
451 * before the timer has a chance to fire.
452 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100453 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
454 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800455 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400456 HRTIMER_MODE_REL_PINNED);
457
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200458 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000459 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200460 /* Enable the perf event */
461 watchdog_nmi_enable(cpu);
462
463 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400464}
465
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000466static void watchdog_disable(unsigned int cpu)
467{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200468 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000469
470 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200471 /*
472 * Disable the perf event first. That prevents that a large delay
473 * between disabling the timer and disabling the perf event causes
474 * the perf NMI to detect a false positive.
475 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000476 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200477 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000478}
479
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200480static void watchdog_cleanup(unsigned int cpu, bool online)
481{
482 watchdog_disable(cpu);
483}
484
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000485static int watchdog_should_run(unsigned int cpu)
486{
487 return __this_cpu_read(hrtimer_interrupts) !=
488 __this_cpu_read(soft_lockup_hrtimer_cnt);
489}
490
491/*
492 * The watchdog thread function - touches the timestamp.
493 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800494 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000495 * default) to reset the softlockup timestamp. If this gets delayed
496 * for more than 2*watchdog_thresh seconds then the debug-printout
497 * triggers in watchdog_timer_fn().
498 */
499static void watchdog(unsigned int cpu)
500{
501 __this_cpu_write(soft_lockup_hrtimer_cnt,
502 __this_cpu_read(hrtimer_interrupts));
503 __touch_watchdog();
504}
Don Zickus58687ac2010-05-07 17:11:44 -0400505
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200506static struct smp_hotplug_thread watchdog_threads = {
507 .store = &softlockup_watchdog,
508 .thread_should_run = watchdog_should_run,
509 .thread_fn = watchdog,
510 .thread_comm = "watchdog/%u",
511 .setup = watchdog_enable,
512 .cleanup = watchdog_cleanup,
513 .park = watchdog_disable,
514 .unpark = watchdog_enable,
515};
516
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200517static void softlockup_update_smpboot_threads(void)
518{
519 lockdep_assert_held(&watchdog_mutex);
520
521 if (!softlockup_threads_initialized)
522 return;
523
524 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
525 &watchdog_allowed_mask);
526 __lockup_detector_cleanup();
527}
528
529/* Temporarily park all watchdog threads */
530static void softlockup_park_all_threads(void)
531{
532 cpumask_clear(&watchdog_allowed_mask);
533 softlockup_update_smpboot_threads();
534}
535
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200536/* Unpark enabled threads */
537static void softlockup_unpark_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200538{
539 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
540 softlockup_update_smpboot_threads();
541}
542
Thomas Gleixner09154982017-09-12 21:37:17 +0200543static void softlockup_reconfigure_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200544{
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200545 watchdog_nmi_reconfigure(false);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200546 softlockup_park_all_threads();
547 set_sample_period();
Thomas Gleixner09154982017-09-12 21:37:17 +0200548 lockup_detector_update_enable();
549 if (watchdog_enabled && watchdog_thresh)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200550 softlockup_unpark_threads();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200551 watchdog_nmi_reconfigure(true);
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200552}
553
554/*
555 * Create the watchdog thread infrastructure.
556 *
557 * The threads are not unparked as watchdog_allowed_mask is empty. When
558 * the threads are sucessfully initialized, take the proper locks and
559 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
560 */
561static __init void softlockup_init_threads(void)
562{
563 int ret;
564
565 /*
566 * If sysctl is off and watchdog got disabled on the command line,
567 * nothing to do here.
568 */
Thomas Gleixner09154982017-09-12 21:37:17 +0200569 lockup_detector_update_enable();
570
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200571 if (!IS_ENABLED(CONFIG_SYSCTL) &&
572 !(watchdog_enabled && watchdog_thresh))
573 return;
574
575 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
576 &watchdog_allowed_mask);
577 if (ret) {
578 pr_err("Failed to initialize soft lockup detector threads\n");
579 return;
580 }
581
582 mutex_lock(&watchdog_mutex);
583 softlockup_threads_initialized = true;
Thomas Gleixner09154982017-09-12 21:37:17 +0200584 softlockup_reconfigure_threads();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200585 mutex_unlock(&watchdog_mutex);
586}
587
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200588#else /* CONFIG_SOFTLOCKUP_DETECTOR */
589static inline int watchdog_park_threads(void) { return 0; }
590static inline void watchdog_unpark_threads(void) { }
591static inline int watchdog_enable_all_cpus(void) { return 0; }
592static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200593static inline void softlockup_init_threads(void) { }
Thomas Gleixner09154982017-09-12 21:37:17 +0200594static void softlockup_reconfigure_threads(void)
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200595{
596 watchdog_nmi_reconfigure(false);
Thomas Gleixner09154982017-09-12 21:37:17 +0200597 lockup_detector_update_enable();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200598 watchdog_nmi_reconfigure(true);
599}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200600#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700601
Thomas Gleixner941154b2017-09-12 21:37:04 +0200602static void __lockup_detector_cleanup(void)
603{
604 lockdep_assert_held(&watchdog_mutex);
605 hardlockup_detector_perf_cleanup();
606}
607
608/**
609 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
610 *
611 * Caller must not hold the cpu hotplug rwsem.
612 */
613void lockup_detector_cleanup(void)
614{
615 mutex_lock(&watchdog_mutex);
616 __lockup_detector_cleanup();
617 mutex_unlock(&watchdog_mutex);
618}
619
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200620/**
621 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
622 *
623 * Special interface for parisc. It prevents lockup detector warnings from
624 * the default pm_poweroff() function which busy loops forever.
625 */
626void lockup_detector_soft_poweroff(void)
627{
628 watchdog_enabled = 0;
629}
630
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800631#ifdef CONFIG_SYSCTL
632
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200633/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200634static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400635{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200636 /* Remove impossible cpus to keep sysctl output clean. */
637 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixner09154982017-09-12 21:37:17 +0200638 softlockup_reconfigure_threads();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700639}
640
641/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700642 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
643 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200644 * caller | table->data points to | 'which'
645 * -------------------|----------------------------|--------------------------
646 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
647 * | | SOFT_WATCHDOG_ENABLED
648 * -------------------|----------------------------|--------------------------
649 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
650 * -------------------|----------------------------|--------------------------
651 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700652 */
653static int proc_watchdog_common(int which, struct ctl_table *table, int write,
654 void __user *buffer, size_t *lenp, loff_t *ppos)
655{
Thomas Gleixner09154982017-09-12 21:37:17 +0200656 int err, old, *param = table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400657
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200658 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200659 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000660
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700661 if (!write) {
Thomas Gleixner09154982017-09-12 21:37:17 +0200662 /*
663 * On read synchronize the userspace interface. This is a
664 * racy snapshot.
665 */
666 *param = (watchdog_enabled & which) != 0;
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700667 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
668 } else {
Thomas Gleixner09154982017-09-12 21:37:17 +0200669 old = READ_ONCE(*param);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700670 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Thomas Gleixner09154982017-09-12 21:37:17 +0200671 if (!err && old != READ_ONCE(*param))
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200672 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700673 }
Thomas Gleixner946d1972017-09-12 21:37:01 +0200674 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200675 cpu_hotplug_enable();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700676 return err;
677}
678
679/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700680 * /proc/sys/kernel/watchdog
681 */
682int proc_watchdog(struct ctl_table *table, int write,
683 void __user *buffer, size_t *lenp, loff_t *ppos)
684{
685 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
686 table, write, buffer, lenp, ppos);
687}
688
689/*
690 * /proc/sys/kernel/nmi_watchdog
691 */
692int proc_nmi_watchdog(struct ctl_table *table, int write,
693 void __user *buffer, size_t *lenp, loff_t *ppos)
694{
695 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
696 table, write, buffer, lenp, ppos);
697}
698
699/*
700 * /proc/sys/kernel/soft_watchdog
701 */
702int proc_soft_watchdog(struct ctl_table *table, int write,
703 void __user *buffer, size_t *lenp, loff_t *ppos)
704{
705 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
706 table, write, buffer, lenp, ppos);
707}
708
709/*
710 * /proc/sys/kernel/watchdog_thresh
711 */
712int proc_watchdog_thresh(struct ctl_table *table, int write,
713 void __user *buffer, size_t *lenp, loff_t *ppos)
714{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200715 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700716
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200717 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200718 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700719
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200720 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200721 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700722
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200723 if (!err && write && old != READ_ONCE(watchdog_thresh))
724 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700725
Thomas Gleixner946d1972017-09-12 21:37:01 +0200726 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200727 cpu_hotplug_enable();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200728 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400729}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700730
731/*
732 * The cpumask is the mask of possible cpus that the watchdog can run
733 * on, not the mask of cpus it is actually running on. This allows the
734 * user to specify a mask that will include cpus that have not yet
735 * been brought online, if desired.
736 */
737int proc_watchdog_cpumask(struct ctl_table *table, int write,
738 void __user *buffer, size_t *lenp, loff_t *ppos)
739{
740 int err;
741
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200742 cpu_hotplug_disable();
Thomas Gleixner946d1972017-09-12 21:37:01 +0200743 mutex_lock(&watchdog_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700744
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700745 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
Thomas Gleixner05ba3de2017-09-12 21:37:08 +0200746 if (!err && write)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200747 proc_watchdog_update();
Thomas Gleixner54901252017-09-12 21:36:59 +0200748
Thomas Gleixner946d1972017-09-12 21:37:01 +0200749 mutex_unlock(&watchdog_mutex);
Thomas Gleixnerb7a34982017-09-12 21:37:00 +0200750 cpu_hotplug_enable();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700751 return err;
752}
Don Zickus58687ac2010-05-07 17:11:44 -0400753#endif /* CONFIG_SYSCTL */
754
Peter Zijlstra004417a2010-11-25 18:38:29 +0100755void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400756{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700757#ifdef CONFIG_NO_HZ_FULL
758 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700759 pr_info("Disabling watchdog on nohz_full cores by default\n");
760 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700761 } else
762 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
763#else
764 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
765#endif
766
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200767 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400768}