blob: 6ad6226535d0e62e549113bd345cc087db023372 [file] [log] [blame]
Don Zickus58687ac2010-05-07 17:11:44 -04001/*
2 * Detect hard and soft lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -05006 * Note: Most of this code is borrowed heavily from the original softlockup
7 * detector, so thanks to Ingo for the initial implementation.
8 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
Don Zickus58687ac2010-05-07 17:11:44 -04009 * to those contributors as well.
10 */
11
Kefeng Wang5f92a7b2017-07-14 14:49:46 -070012#define pr_fmt(fmt) "watchdog: " fmt
Andrew Morton45019802012-03-23 15:01:55 -070013
Don Zickus58687ac2010-05-07 17:11:44 -040014#include <linux/mm.h>
15#include <linux/cpu.h>
16#include <linux/nmi.h>
17#include <linux/init.h>
Don Zickus58687ac2010-05-07 17:11:44 -040018#include <linux/module.h>
19#include <linux/sysctl.h>
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +000020#include <linux/smpboot.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060021#include <linux/sched/rt.h>
Ingo Molnarae7e81c2017-02-01 18:07:51 +010022#include <uapi/linux/sched/types.h>
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070023#include <linux/tick.h>
Tejun Heo82607adc2015-12-08 11:28:04 -050024#include <linux/workqueue.h>
Ingo Molnare6017572017-02-01 16:36:40 +010025#include <linux/sched/clock.h>
Ingo Molnarb17b0152017-02-08 18:51:35 +010026#include <linux/sched/debug.h>
Don Zickus58687ac2010-05-07 17:11:44 -040027
28#include <asm/irq_regs.h>
Eric B Munson5d1c0f42012-03-10 14:37:28 -050029#include <linux/kvm_para.h>
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -070030#include <linux/kthread.h>
Don Zickus58687ac2010-05-07 17:11:44 -040031
Thomas Gleixner946d1972017-09-12 21:37:01 +020032static DEFINE_MUTEX(watchdog_mutex);
Peter Zijlstraab992dc2015-05-18 11:31:50 +020033
Nicholas Piggin05a4a952017-07-12 14:35:46 -070034#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
Thomas Gleixner09154982017-09-12 21:37:17 +020035# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
36# define NMI_WATCHDOG_DEFAULT 1
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070037#else
Thomas Gleixner09154982017-09-12 21:37:17 +020038# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
39# define NMI_WATCHDOG_DEFAULT 0
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070040#endif
Nicholas Piggin05a4a952017-07-12 14:35:46 -070041
Thomas Gleixner09154982017-09-12 21:37:17 +020042unsigned long __read_mostly watchdog_enabled;
43int __read_mostly watchdog_user_enabled = 1;
44int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
45int __read_mostly soft_watchdog_user_enabled = 1;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020046int __read_mostly watchdog_thresh = 10;
Thomas Gleixnera994a312017-09-12 21:37:19 +020047int __read_mostly nmi_watchdog_available;
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +020048
49struct cpumask watchdog_allowed_mask __read_mostly;
50static bool softlockup_threads_initialized __read_mostly;
51
52struct cpumask watchdog_cpumask __read_mostly;
53unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
54
Nicholas Piggin05a4a952017-07-12 14:35:46 -070055#ifdef CONFIG_HARDLOCKUP_DETECTOR
Nicholas Piggin05a4a952017-07-12 14:35:46 -070056/*
57 * Should we panic when a soft-lockup or hard-lockup occurs:
58 */
59unsigned int __read_mostly hardlockup_panic =
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
61/*
62 * We may not want to enable hard lockup detection by default in all cases,
63 * for example when running the kernel as a guest on a hypervisor. In these
64 * cases this function can be called to disable hard lockup detection. This
65 * function should only be executed once by the boot processor before the
66 * kernel command line parameters are parsed, because otherwise it is not
67 * possible to override this in hardlockup_panic_setup().
68 */
Thomas Gleixner7a355822017-09-12 21:37:02 +020069void __init hardlockup_detector_disable(void)
Nicholas Piggin05a4a952017-07-12 14:35:46 -070070{
Thomas Gleixner09154982017-09-12 21:37:17 +020071 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070072}
73
74static int __init hardlockup_panic_setup(char *str)
75{
76 if (!strncmp(str, "panic", 5))
77 hardlockup_panic = 1;
78 else if (!strncmp(str, "nopanic", 7))
79 hardlockup_panic = 0;
80 else if (!strncmp(str, "0", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020081 nmi_watchdog_user_enabled = 0;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070082 else if (!strncmp(str, "1", 1))
Thomas Gleixner09154982017-09-12 21:37:17 +020083 nmi_watchdog_user_enabled = 1;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070084 return 1;
85}
86__setup("nmi_watchdog=", hardlockup_panic_setup);
87
Thomas Gleixner368a7e22017-09-12 21:37:07 +020088# ifdef CONFIG_SMP
89int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Nicholas Piggin05a4a952017-07-12 14:35:46 -070090
Thomas Gleixner368a7e22017-09-12 21:37:07 +020091static int __init hardlockup_all_cpu_backtrace_setup(char *str)
92{
93 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
94 return 1;
95}
96__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
97# endif /* CONFIG_SMP */
98#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -070099
Ulrich Obergfellec6a9062015-09-04 15:45:28 -0700100/*
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700101 * These functions can be overridden if an architecture implements its
102 * own hardlockup detector.
Nicholas Piggina10a8422017-07-12 14:35:49 -0700103 *
104 * watchdog_nmi_enable/disable can be implemented to start and stop when
105 * softlockup watchdog threads start and stop. The arch must select the
106 * SOFTLOCKUP_DETECTOR Kconfig.
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700107 */
108int __weak watchdog_nmi_enable(unsigned int cpu)
109{
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200110 hardlockup_detector_perf_enable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700111 return 0;
112}
Thomas Gleixner941154b2017-09-12 21:37:04 +0200113
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700114void __weak watchdog_nmi_disable(unsigned int cpu)
115{
Thomas Gleixner941154b2017-09-12 21:37:04 +0200116 hardlockup_detector_perf_disable();
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700117}
118
Thomas Gleixnera994a312017-09-12 21:37:19 +0200119/* Return 0, if a NMI watchdog is available. Error code otherwise */
120int __weak __init watchdog_nmi_probe(void)
121{
122 return hardlockup_detector_perf_init();
123}
124
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200125/**
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200126 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200127 *
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200128 * The reconfiguration steps are:
129 * watchdog_nmi_stop();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200130 * update_variables();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200131 * watchdog_nmi_start();
132 */
133void __weak watchdog_nmi_stop(void) { }
134
135/**
136 * watchdog_nmi_start - Start the watchdog after reconfiguration
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200137 *
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200138 * Counterpart to watchdog_nmi_stop().
139 *
140 * The following variables have been updated in update_variables() and
141 * contain the currently valid configuration:
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200142 * - watchdog_enabled
Nicholas Piggina10a8422017-07-12 14:35:49 -0700143 * - watchdog_thresh
144 * - watchdog_cpumask
Nicholas Piggina10a8422017-07-12 14:35:49 -0700145 */
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200146void __weak watchdog_nmi_start(void) { }
Nicholas Piggina10a8422017-07-12 14:35:49 -0700147
Thomas Gleixner09154982017-09-12 21:37:17 +0200148/**
149 * lockup_detector_update_enable - Update the sysctl enable bit
150 *
151 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
152 * can't race with watchdog_nmi_disable().
153 */
154static void lockup_detector_update_enable(void)
155{
156 watchdog_enabled = 0;
157 if (!watchdog_user_enabled)
158 return;
Thomas Gleixnera994a312017-09-12 21:37:19 +0200159 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
Thomas Gleixner09154982017-09-12 21:37:17 +0200160 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
161 if (soft_watchdog_user_enabled)
162 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
163}
164
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700165#ifdef CONFIG_SOFTLOCKUP_DETECTOR
166
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200167/* Global variables, exported for sysctl */
168unsigned int __read_mostly softlockup_panic =
169 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200170
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800171static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -0400172
173static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
174static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
175static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
176static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -0400177static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000178static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
179static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700180static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Don Zickus58687ac2010-05-07 17:11:44 -0400181static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
Aaron Tomlined235872014-06-23 13:22:05 -0700182static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400183
Don Zickus58687ac2010-05-07 17:11:44 -0400184static int __init softlockup_panic_setup(char *str)
185{
186 softlockup_panic = simple_strtoul(str, NULL, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400187 return 1;
188}
189__setup("softlockup_panic=", softlockup_panic_setup);
190
191static int __init nowatchdog_setup(char *str)
192{
Thomas Gleixner09154982017-09-12 21:37:17 +0200193 watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400194 return 1;
195}
196__setup("nowatchdog", nowatchdog_setup);
197
Don Zickus58687ac2010-05-07 17:11:44 -0400198static int __init nosoftlockup_setup(char *str)
199{
Thomas Gleixner09154982017-09-12 21:37:17 +0200200 soft_watchdog_user_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400201 return 1;
202}
203__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700204
Aaron Tomlined235872014-06-23 13:22:05 -0700205#ifdef CONFIG_SMP
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200206int __read_mostly sysctl_softlockup_all_cpu_backtrace;
207
Aaron Tomlined235872014-06-23 13:22:05 -0700208static int __init softlockup_all_cpu_backtrace_setup(char *str)
209{
Thomas Gleixner368a7e22017-09-12 21:37:07 +0200210 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
Aaron Tomlined235872014-06-23 13:22:05 -0700211 return 1;
212}
213__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700214#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400215
Thomas Gleixner941154b2017-09-12 21:37:04 +0200216static void __lockup_detector_cleanup(void);
217
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700218/*
219 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
220 * lockups can have false positives under extreme conditions. So we generally
221 * want a higher threshold for soft lockups than for hard lockups. So we couple
222 * the thresholds with a factor: we make the soft threshold twice the amount of
223 * time the hard threshold is.
224 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200225static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700226{
227 return watchdog_thresh * 2;
228}
Don Zickus58687ac2010-05-07 17:11:44 -0400229
230/*
231 * Returns seconds, approximately. We don't need nanosecond
232 * resolution, and we don't need to waste time with a big divide when
233 * 2^30ns == 1.074s.
234 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900235static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400236{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800237 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400238}
239
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800240static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400241{
242 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700243 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500244 * the divide by 5 is to give hrtimer several chances (two
245 * or three with the current relation between the soft
246 * and hard thresholds) to increment before the
247 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400248 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800249 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Thomas Gleixner7edaeb62017-08-15 09:50:13 +0200250 watchdog_update_hrtimer_threshold(sample_period);
Don Zickus58687ac2010-05-07 17:11:44 -0400251}
252
253/* Commands for resetting the watchdog */
254static void __touch_watchdog(void)
255{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900256 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400257}
258
Tejun Heo03e0d462015-12-08 11:28:04 -0500259/**
260 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
261 *
262 * Call when the scheduler may have stalled for legitimate reasons
263 * preventing the watchdog task from executing - e.g. the scheduler
264 * entering idle state. This should only be used for scheduler events.
265 * Use touch_softlockup_watchdog() for everything else.
266 */
267void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400268{
Andrew Morton78611442014-04-18 15:07:12 -0700269 /*
270 * Preemption can be enabled. It doesn't matter which CPU's timestamp
271 * gets zeroed here, so use the raw_ operation.
272 */
273 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400274}
Tejun Heo03e0d462015-12-08 11:28:04 -0500275
276void touch_softlockup_watchdog(void)
277{
278 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500279 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500280}
Ingo Molnar0167c782010-05-13 08:53:33 +0200281EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400282
Don Zickus332fbdb2010-05-07 17:11:45 -0400283void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400284{
285 int cpu;
286
287 /*
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200288 * watchdog_mutex cannpt be taken here, as this might be called
289 * from (soft)interrupt context, so the access to
290 * watchdog_allowed_cpumask might race with a concurrent update.
291 *
292 * The watchdog time stamp can race against a concurrent real
293 * update as well, the only side effect might be a cycle delay for
294 * the softlockup check.
Don Zickus58687ac2010-05-07 17:11:44 -0400295 */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200296 for_each_cpu(cpu, &watchdog_allowed_mask)
Don Zickus58687ac2010-05-07 17:11:44 -0400297 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500298 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400299}
300
Don Zickus58687ac2010-05-07 17:11:44 -0400301void touch_softlockup_watchdog_sync(void)
302{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500303 __this_cpu_write(softlockup_touch_sync, true);
304 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400305}
306
Don Zickus26e09c62010-05-17 18:06:04 -0400307static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400308{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900309 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400310
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800311 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700312 /* Warn about unreasonable delays. */
313 if (time_after(now, touch_ts + get_softlockup_thresh()))
314 return now - touch_ts;
315 }
Don Zickus58687ac2010-05-07 17:11:44 -0400316 return 0;
317}
318
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700319/* watchdog detector functions */
320bool is_hardlockup(void)
321{
322 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
323
324 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
325 return true;
326
327 __this_cpu_write(hrtimer_interrupts_saved, hrint);
328 return false;
329}
330
Don Zickus58687ac2010-05-07 17:11:44 -0400331static void watchdog_interrupt_count(void)
332{
Christoph Lameter909ea962010-12-08 16:22:55 +0100333 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400334}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000335
Don Zickus58687ac2010-05-07 17:11:44 -0400336/* watchdog kicker functions */
337static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
338{
Christoph Lameter909ea962010-12-08 16:22:55 +0100339 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400340 struct pt_regs *regs = get_irq_regs();
341 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700342 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400343
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200344 if (!watchdog_enabled)
Don Zickusb94f5112017-01-24 15:17:53 -0800345 return HRTIMER_NORESTART;
346
Don Zickus58687ac2010-05-07 17:11:44 -0400347 /* kick the hardlockup detector */
348 watchdog_interrupt_count();
349
350 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100351 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400352
353 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800354 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400355
356 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100357 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400358 /*
359 * If the time stamp was touched atomically
360 * make sure the scheduler tick is up to date.
361 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100362 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400363 sched_clock_tick();
364 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500365
366 /* Clear the guest paused flag on watchdog reset */
367 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400368 __touch_watchdog();
369 return HRTIMER_RESTART;
370 }
371
372 /* check for a softlockup
373 * This is done by making sure a high priority task is
374 * being scheduled. The task touches the watchdog to
375 * indicate it is getting cpu time. If it hasn't then
376 * this is a good indication some task is hogging the cpu
377 */
Don Zickus26e09c62010-05-17 18:06:04 -0400378 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400379 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500380 /*
381 * If a virtual machine is stopped by the host it can look to
382 * the watchdog like a soft lockup, check to see if the host
383 * stopped the vm before we issue the warning
384 */
385 if (kvm_check_and_clear_guest_paused())
386 return HRTIMER_RESTART;
387
Don Zickus58687ac2010-05-07 17:11:44 -0400388 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700389 if (__this_cpu_read(soft_watchdog_warn) == true) {
390 /*
391 * When multiple processes are causing softlockups the
392 * softlockup detector only warns on the first one
393 * because the code relies on a full quiet cycle to
394 * re-arm. The second process prevents the quiet cycle
395 * and never gets reported. Use task pointers to detect
396 * this.
397 */
398 if (__this_cpu_read(softlockup_task_ptr_saved) !=
399 current) {
400 __this_cpu_write(soft_watchdog_warn, false);
401 __touch_watchdog();
402 }
Don Zickus58687ac2010-05-07 17:11:44 -0400403 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700404 }
Don Zickus58687ac2010-05-07 17:11:44 -0400405
Aaron Tomlined235872014-06-23 13:22:05 -0700406 if (softlockup_all_cpu_backtrace) {
407 /* Prevent multiple soft-lockup reports if one cpu is already
408 * engaged in dumping cpu back traces
409 */
410 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
411 /* Someone else will report us. Let's give up */
412 __this_cpu_write(soft_watchdog_warn, true);
413 return HRTIMER_RESTART;
414 }
415 }
416
Fabian Frederick656c3b72014-08-06 16:04:03 -0700417 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400418 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400419 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700420 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400421 print_modules();
422 print_irqtrace_events(current);
423 if (regs)
424 show_regs(regs);
425 else
426 dump_stack();
427
Aaron Tomlined235872014-06-23 13:22:05 -0700428 if (softlockup_all_cpu_backtrace) {
429 /* Avoid generating two back traces for current
430 * given that one is already made above
431 */
432 trigger_allbutself_cpu_backtrace();
433
434 clear_bit(0, &soft_lockup_nmi_warn);
435 /* Barrier to sync with other cpus */
436 smp_mb__after_atomic();
437 }
438
Josh Hunt69361ee2014-08-08 14:22:31 -0700439 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400440 if (softlockup_panic)
441 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100442 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400443 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100444 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400445
446 return HRTIMER_RESTART;
447}
448
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000449static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400450{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000451 struct sched_param param = { .sched_priority = prio };
452
453 sched_setscheduler(current, policy, &param);
454}
455
456static void watchdog_enable(unsigned int cpu)
457{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200458 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400459
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200460 /*
461 * Start the timer first to prevent the NMI watchdog triggering
462 * before the timer has a chance to fire.
463 */
Bjørn Mork3935e8952012-12-19 20:51:31 +0100464 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
465 hrtimer->function = watchdog_timer_fn;
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800466 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400467 HRTIMER_MODE_REL_PINNED);
468
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200469 /* Initialize timestamp */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000470 __touch_watchdog();
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200471 /* Enable the perf event */
Thomas Gleixner146c9d02017-09-12 21:37:21 +0200472 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
473 watchdog_nmi_enable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200474
475 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
Don Zickus58687ac2010-05-07 17:11:44 -0400476}
477
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000478static void watchdog_disable(unsigned int cpu)
479{
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200480 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000481
482 watchdog_set_prio(SCHED_NORMAL, 0);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200483 /*
484 * Disable the perf event first. That prevents that a large delay
485 * between disabling the timer and disabling the perf event causes
486 * the perf NMI to detect a false positive.
487 */
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000488 watchdog_nmi_disable(cpu);
Thomas Gleixner01f0a022017-09-12 21:37:05 +0200489 hrtimer_cancel(hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000490}
491
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200492static void watchdog_cleanup(unsigned int cpu, bool online)
493{
494 watchdog_disable(cpu);
495}
496
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000497static int watchdog_should_run(unsigned int cpu)
498{
499 return __this_cpu_read(hrtimer_interrupts) !=
500 __this_cpu_read(soft_lockup_hrtimer_cnt);
501}
502
503/*
504 * The watchdog thread function - touches the timestamp.
505 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800506 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000507 * default) to reset the softlockup timestamp. If this gets delayed
508 * for more than 2*watchdog_thresh seconds then the debug-printout
509 * triggers in watchdog_timer_fn().
510 */
511static void watchdog(unsigned int cpu)
512{
513 __this_cpu_write(soft_lockup_hrtimer_cnt,
514 __this_cpu_read(hrtimer_interrupts));
515 __touch_watchdog();
516}
Don Zickus58687ac2010-05-07 17:11:44 -0400517
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200518static struct smp_hotplug_thread watchdog_threads = {
519 .store = &softlockup_watchdog,
520 .thread_should_run = watchdog_should_run,
521 .thread_fn = watchdog,
522 .thread_comm = "watchdog/%u",
523 .setup = watchdog_enable,
524 .cleanup = watchdog_cleanup,
525 .park = watchdog_disable,
526 .unpark = watchdog_enable,
527};
528
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200529static void softlockup_update_smpboot_threads(void)
530{
531 lockdep_assert_held(&watchdog_mutex);
532
533 if (!softlockup_threads_initialized)
534 return;
535
536 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
537 &watchdog_allowed_mask);
538 __lockup_detector_cleanup();
539}
540
541/* Temporarily park all watchdog threads */
542static void softlockup_park_all_threads(void)
543{
544 cpumask_clear(&watchdog_allowed_mask);
545 softlockup_update_smpboot_threads();
546}
547
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200548/* Unpark enabled threads */
549static void softlockup_unpark_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200550{
551 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
552 softlockup_update_smpboot_threads();
553}
554
Thomas Gleixner09154982017-09-12 21:37:17 +0200555static void softlockup_reconfigure_threads(void)
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200556{
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200557 watchdog_nmi_stop();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200558 softlockup_park_all_threads();
559 set_sample_period();
Thomas Gleixner09154982017-09-12 21:37:17 +0200560 lockup_detector_update_enable();
561 if (watchdog_enabled && watchdog_thresh)
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200562 softlockup_unpark_threads();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200563 watchdog_nmi_start();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200564}
565
566/*
567 * Create the watchdog thread infrastructure.
568 *
569 * The threads are not unparked as watchdog_allowed_mask is empty. When
570 * the threads are sucessfully initialized, take the proper locks and
571 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
572 */
573static __init void softlockup_init_threads(void)
574{
575 int ret;
576
577 /*
578 * If sysctl is off and watchdog got disabled on the command line,
579 * nothing to do here.
580 */
Thomas Gleixner09154982017-09-12 21:37:17 +0200581 lockup_detector_update_enable();
582
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200583 if (!IS_ENABLED(CONFIG_SYSCTL) &&
584 !(watchdog_enabled && watchdog_thresh))
585 return;
586
587 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
588 &watchdog_allowed_mask);
589 if (ret) {
590 pr_err("Failed to initialize soft lockup detector threads\n");
591 return;
592 }
593
594 mutex_lock(&watchdog_mutex);
595 softlockup_threads_initialized = true;
Thomas Gleixner09154982017-09-12 21:37:17 +0200596 softlockup_reconfigure_threads();
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200597 mutex_unlock(&watchdog_mutex);
598}
599
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200600#else /* CONFIG_SOFTLOCKUP_DETECTOR */
601static inline int watchdog_park_threads(void) { return 0; }
602static inline void watchdog_unpark_threads(void) { }
603static inline int watchdog_enable_all_cpus(void) { return 0; }
604static inline void watchdog_disable_all_cpus(void) { }
Thomas Gleixner2eb25272017-09-12 21:37:10 +0200605static inline void softlockup_init_threads(void) { }
Thomas Gleixner09154982017-09-12 21:37:17 +0200606static void softlockup_reconfigure_threads(void)
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200607{
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200608 watchdog_nmi_stop();
Thomas Gleixner09154982017-09-12 21:37:17 +0200609 lockup_detector_update_enable();
Thomas Gleixner6b9dc482017-10-02 12:34:50 +0200610 watchdog_nmi_start();
Thomas Gleixner6592ad22017-09-12 21:37:16 +0200611}
Thomas Gleixner2b9d7f22017-09-12 21:37:06 +0200612#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
Nicholas Piggin05a4a952017-07-12 14:35:46 -0700613
Thomas Gleixner941154b2017-09-12 21:37:04 +0200614static void __lockup_detector_cleanup(void)
615{
616 lockdep_assert_held(&watchdog_mutex);
617 hardlockup_detector_perf_cleanup();
618}
619
620/**
621 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
622 *
623 * Caller must not hold the cpu hotplug rwsem.
624 */
625void lockup_detector_cleanup(void)
626{
627 mutex_lock(&watchdog_mutex);
628 __lockup_detector_cleanup();
629 mutex_unlock(&watchdog_mutex);
630}
631
Thomas Gleixner6554fd82017-09-12 21:36:57 +0200632/**
633 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
634 *
635 * Special interface for parisc. It prevents lockup detector warnings from
636 * the default pm_poweroff() function which busy loops forever.
637 */
638void lockup_detector_soft_poweroff(void)
639{
640 watchdog_enabled = 0;
641}
642
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800643#ifdef CONFIG_SYSCTL
644
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200645/* Propagate any changes to the watchdog threads */
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200646static void proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400647{
Thomas Gleixnere8b62b22017-09-12 21:37:12 +0200648 /* Remove impossible cpus to keep sysctl output clean. */
649 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
Thomas Gleixner09154982017-09-12 21:37:17 +0200650 softlockup_reconfigure_threads();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700651}
652
653/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700654 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
655 *
Thomas Gleixner7feeb9c2017-09-12 21:37:15 +0200656 * caller | table->data points to | 'which'
657 * -------------------|----------------------------|--------------------------
658 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
659 * | | SOFT_WATCHDOG_ENABLED
660 * -------------------|----------------------------|--------------------------
661 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
662 * -------------------|----------------------------|--------------------------
663 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700664 */
665static int proc_watchdog_common(int which, struct ctl_table *table, int write,
666 void __user *buffer, size_t *lenp, loff_t *ppos)
667{
Thomas Gleixner09154982017-09-12 21:37:17 +0200668 int err, old, *param = table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400669
Thomas Gleixner946d1972017-09-12 21:37:01 +0200670 mutex_lock(&watchdog_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000671
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700672 if (!write) {
Thomas Gleixner09154982017-09-12 21:37:17 +0200673 /*
674 * On read synchronize the userspace interface. This is a
675 * racy snapshot.
676 */
677 *param = (watchdog_enabled & which) != 0;
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700678 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
679 } else {
Thomas Gleixner09154982017-09-12 21:37:17 +0200680 old = READ_ONCE(*param);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700681 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Thomas Gleixner09154982017-09-12 21:37:17 +0200682 if (!err && old != READ_ONCE(*param))
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200683 proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700684 }
Thomas Gleixner946d1972017-09-12 21:37:01 +0200685 mutex_unlock(&watchdog_mutex);
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700686 return err;
687}
688
689/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700690 * /proc/sys/kernel/watchdog
691 */
692int proc_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|SOFT_WATCHDOG_ENABLED,
696 table, write, buffer, lenp, ppos);
697}
698
699/*
700 * /proc/sys/kernel/nmi_watchdog
701 */
702int proc_nmi_watchdog(struct ctl_table *table, int write,
703 void __user *buffer, size_t *lenp, loff_t *ppos)
704{
Thomas Gleixnera994a312017-09-12 21:37:19 +0200705 if (!nmi_watchdog_available && write)
706 return -ENOTSUPP;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700707 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
708 table, write, buffer, lenp, ppos);
709}
710
711/*
712 * /proc/sys/kernel/soft_watchdog
713 */
714int proc_soft_watchdog(struct ctl_table *table, int write,
715 void __user *buffer, size_t *lenp, loff_t *ppos)
716{
717 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
718 table, write, buffer, lenp, ppos);
719}
720
721/*
722 * /proc/sys/kernel/watchdog_thresh
723 */
724int proc_watchdog_thresh(struct ctl_table *table, int write,
725 void __user *buffer, size_t *lenp, loff_t *ppos)
726{
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200727 int err, old;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700728
Thomas Gleixner946d1972017-09-12 21:37:01 +0200729 mutex_lock(&watchdog_mutex);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700730
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200731 old = READ_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200732 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700733
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200734 if (!err && write && old != READ_ONCE(watchdog_thresh))
735 proc_watchdog_update();
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700736
Thomas Gleixner946d1972017-09-12 21:37:01 +0200737 mutex_unlock(&watchdog_mutex);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200738 return err;
Don Zickus58687ac2010-05-07 17:11:44 -0400739}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700740
741/*
742 * The cpumask is the mask of possible cpus that the watchdog can run
743 * on, not the mask of cpus it is actually running on. This allows the
744 * user to specify a mask that will include cpus that have not yet
745 * been brought online, if desired.
746 */
747int proc_watchdog_cpumask(struct ctl_table *table, int write,
748 void __user *buffer, size_t *lenp, loff_t *ppos)
749{
750 int err;
751
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);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700759 return err;
760}
Don Zickus58687ac2010-05-07 17:11:44 -0400761#endif /* CONFIG_SYSCTL */
762
Peter Zijlstra004417a2010-11-25 18:38:29 +0100763void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400764{
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700765#ifdef CONFIG_NO_HZ_FULL
766 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -0700767 pr_info("Disabling watchdog on nohz_full cores by default\n");
768 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700769 } else
770 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
771#else
772 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
773#endif
774
Thomas Gleixnera994a312017-09-12 21:37:19 +0200775 if (!watchdog_nmi_probe())
776 nmi_watchdog_available = true;
Thomas Gleixnerd57108d2017-09-12 21:37:11 +0200777 softlockup_init_threads();
Don Zickus58687ac2010-05-07 17:11:44 -0400778}