blob: 9acb29f280ec95813612a913307e4dcaec8b7e03 [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
Andrew Morton45019802012-03-23 15:01:55 -070012#define pr_fmt(fmt) "NMI watchdog: " fmt
13
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>
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070022#include <linux/tick.h>
Tejun Heo82607adc2015-12-08 11:28:04 -050023#include <linux/workqueue.h>
Don Zickus58687ac2010-05-07 17:11:44 -040024
25#include <asm/irq_regs.h>
Eric B Munson5d1c0f42012-03-10 14:37:28 -050026#include <linux/kvm_para.h>
Don Zickus58687ac2010-05-07 17:11:44 -040027#include <linux/perf_event.h>
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -070028#include <linux/kthread.h>
Don Zickus58687ac2010-05-07 17:11:44 -040029
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070030/*
31 * The run state of the lockup detectors is controlled by the content of the
32 * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
33 * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
34 *
35 * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
36 * are variables that are only used as an 'interface' between the parameters
37 * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
38 * 'watchdog_thresh' variable is handled differently because its value is not
39 * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
40 * is equal zero.
41 */
42#define NMI_WATCHDOG_ENABLED_BIT 0
43#define SOFT_WATCHDOG_ENABLED_BIT 1
44#define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT)
45#define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT)
46
Peter Zijlstraab992dc2015-05-18 11:31:50 +020047static DEFINE_MUTEX(watchdog_proc_mutex);
48
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070049#ifdef CONFIG_HARDLOCKUP_DETECTOR
50static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
51#else
52static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
53#endif
54int __read_mostly nmi_watchdog_enabled;
55int __read_mostly soft_watchdog_enabled;
56int __read_mostly watchdog_user_enabled;
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -070057int __read_mostly watchdog_thresh = 10;
Ulrich Obergfell84d56e62015-04-14 15:43:55 -070058
Aaron Tomlined235872014-06-23 13:22:05 -070059#ifdef CONFIG_SMP
60int __read_mostly sysctl_softlockup_all_cpu_backtrace;
Jiri Kosina55537872015-11-05 18:44:41 -080061int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
Aaron Tomlined235872014-06-23 13:22:05 -070062#else
63#define sysctl_softlockup_all_cpu_backtrace 0
Jiri Kosina55537872015-11-05 18:44:41 -080064#define sysctl_hardlockup_all_cpu_backtrace 0
Aaron Tomlined235872014-06-23 13:22:05 -070065#endif
Chris Metcalffe4ba3c2015-06-24 16:55:45 -070066static struct cpumask watchdog_cpumask __read_mostly;
67unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
68
69/* Helper for online, unparked cpus. */
70#define for_each_watchdog_cpu(cpu) \
71 for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
Aaron Tomlined235872014-06-23 13:22:05 -070072
Ulrich Obergfellec6a9062015-09-04 15:45:28 -070073/*
74 * The 'watchdog_running' variable is set to 1 when the watchdog threads
75 * are registered/started and is set to 0 when the watchdog threads are
76 * unregistered/stopped, so it is an indicator whether the threads exist.
77 */
Frederic Weisbecker3c00ea82013-05-19 20:45:15 +020078static int __read_mostly watchdog_running;
Ulrich Obergfellec6a9062015-09-04 15:45:28 -070079/*
80 * If a subsystem has a need to deactivate the watchdog temporarily, it
81 * can use the suspend/resume interface to achieve this. The content of
82 * the 'watchdog_suspended' variable reflects this state. Existing threads
83 * are parked/unparked by the lockup_detector_{suspend|resume} functions
84 * (see comment blocks pertaining to those functions for further details).
85 *
86 * 'watchdog_suspended' also prevents threads from being registered/started
87 * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
88 * of 'watchdog_running' cannot change while the watchdog is deactivated
89 * temporarily (see related code in 'proc' handlers).
90 */
91static int __read_mostly watchdog_suspended;
92
Chuansheng Liu0f34c402012-12-17 15:59:50 -080093static u64 __read_mostly sample_period;
Don Zickus58687ac2010-05-07 17:11:44 -040094
95static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
96static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
97static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
98static DEFINE_PER_CPU(bool, softlockup_touch_sync);
Don Zickus58687ac2010-05-07 17:11:44 -040099static DEFINE_PER_CPU(bool, soft_watchdog_warn);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000100static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
101static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
chai wenb1a8de12014-10-09 15:25:17 -0700102static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200103#ifdef CONFIG_HARDLOCKUP_DETECTOR
Don Zickuscafcd802010-05-14 11:11:21 -0400104static DEFINE_PER_CPU(bool, hard_watchdog_warn);
105static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
Don Zickus58687ac2010-05-07 17:11:44 -0400106static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
107static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
108#endif
Aaron Tomlined235872014-06-23 13:22:05 -0700109static unsigned long soft_lockup_nmi_warn;
Don Zickus58687ac2010-05-07 17:11:44 -0400110
Don Zickus58687ac2010-05-07 17:11:44 -0400111/* boot commands */
112/*
113 * Should we panic when a soft-lockup or hard-lockup occurs:
114 */
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200115#ifdef CONFIG_HARDLOCKUP_DETECTOR
Don Zickusac1f5912015-11-05 18:44:44 -0800116unsigned int __read_mostly hardlockup_panic =
Don Zickusfef2c9b2011-03-22 16:34:16 -0700117 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
Jiri Kosina55537872015-11-05 18:44:41 -0800118static unsigned long hardlockup_allcpu_dumped;
Ulrich Obergfell6e7458a2014-10-13 15:55:35 -0700119/*
120 * We may not want to enable hard lockup detection by default in all cases,
121 * for example when running the kernel as a guest on a hypervisor. In these
122 * cases this function can be called to disable hard lockup detection. This
123 * function should only be executed once by the boot processor before the
124 * kernel command line parameters are parsed, because otherwise it is not
125 * possible to override this in hardlockup_panic_setup().
126 */
Ulrich Obergfell692297d2015-04-14 15:44:19 -0700127void hardlockup_detector_disable(void)
Ulrich Obergfell6e7458a2014-10-13 15:55:35 -0700128{
Ulrich Obergfell692297d2015-04-14 15:44:19 -0700129 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
Ulrich Obergfell6e7458a2014-10-13 15:55:35 -0700130}
131
Don Zickus58687ac2010-05-07 17:11:44 -0400132static int __init hardlockup_panic_setup(char *str)
133{
134 if (!strncmp(str, "panic", 5))
135 hardlockup_panic = 1;
Don Zickusfef2c9b2011-03-22 16:34:16 -0700136 else if (!strncmp(str, "nopanic", 7))
137 hardlockup_panic = 0;
Don Zickus5dc30552010-11-29 17:07:17 -0500138 else if (!strncmp(str, "0", 1))
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700139 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
140 else if (!strncmp(str, "1", 1))
141 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
Don Zickus58687ac2010-05-07 17:11:44 -0400142 return 1;
143}
144__setup("nmi_watchdog=", hardlockup_panic_setup);
145#endif
146
147unsigned int __read_mostly softlockup_panic =
148 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
149
150static int __init softlockup_panic_setup(char *str)
151{
152 softlockup_panic = simple_strtoul(str, NULL, 0);
153
154 return 1;
155}
156__setup("softlockup_panic=", softlockup_panic_setup);
157
158static int __init nowatchdog_setup(char *str)
159{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700160 watchdog_enabled = 0;
Don Zickus58687ac2010-05-07 17:11:44 -0400161 return 1;
162}
163__setup("nowatchdog", nowatchdog_setup);
164
Don Zickus58687ac2010-05-07 17:11:44 -0400165static int __init nosoftlockup_setup(char *str)
166{
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700167 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
Don Zickus58687ac2010-05-07 17:11:44 -0400168 return 1;
169}
170__setup("nosoftlockup", nosoftlockup_setup);
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700171
Aaron Tomlined235872014-06-23 13:22:05 -0700172#ifdef CONFIG_SMP
173static int __init softlockup_all_cpu_backtrace_setup(char *str)
174{
175 sysctl_softlockup_all_cpu_backtrace =
176 !!simple_strtol(str, NULL, 0);
177 return 1;
178}
179__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
Jiri Kosina55537872015-11-05 18:44:41 -0800180static int __init hardlockup_all_cpu_backtrace_setup(char *str)
181{
182 sysctl_hardlockup_all_cpu_backtrace =
183 !!simple_strtol(str, NULL, 0);
184 return 1;
185}
186__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
Aaron Tomlined235872014-06-23 13:22:05 -0700187#endif
Don Zickus58687ac2010-05-07 17:11:44 -0400188
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700189/*
190 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
191 * lockups can have false positives under extreme conditions. So we generally
192 * want a higher threshold for soft lockups than for hard lockups. So we couple
193 * the thresholds with a factor: we make the soft threshold twice the amount of
194 * time the hard threshold is.
195 */
Ingo Molnar6e9101a2011-05-24 05:43:18 +0200196static int get_softlockup_thresh(void)
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700197{
198 return watchdog_thresh * 2;
199}
Don Zickus58687ac2010-05-07 17:11:44 -0400200
201/*
202 * Returns seconds, approximately. We don't need nanosecond
203 * resolution, and we don't need to waste time with a big divide when
204 * 2^30ns == 1.074s.
205 */
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900206static unsigned long get_timestamp(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400207{
Cyril Bur545a2bf2015-02-12 15:01:24 -0800208 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
Don Zickus58687ac2010-05-07 17:11:44 -0400209}
210
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800211static void set_sample_period(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400212{
213 /*
Mandeep Singh Baines586692a2011-05-22 22:10:22 -0700214 * convert watchdog_thresh from seconds to ns
Fernando Luis Vázquez Cao86f5e6a2012-02-09 17:42:22 -0500215 * the divide by 5 is to give hrtimer several chances (two
216 * or three with the current relation between the soft
217 * and hard thresholds) to increment before the
218 * hardlockup detector generates a warning
Don Zickus58687ac2010-05-07 17:11:44 -0400219 */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800220 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
Don Zickus58687ac2010-05-07 17:11:44 -0400221}
222
223/* Commands for resetting the watchdog */
224static void __touch_watchdog(void)
225{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900226 __this_cpu_write(watchdog_touch_ts, get_timestamp());
Don Zickus58687ac2010-05-07 17:11:44 -0400227}
228
Tejun Heo03e0d462015-12-08 11:28:04 -0500229/**
230 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
231 *
232 * Call when the scheduler may have stalled for legitimate reasons
233 * preventing the watchdog task from executing - e.g. the scheduler
234 * entering idle state. This should only be used for scheduler events.
235 * Use touch_softlockup_watchdog() for everything else.
236 */
237void touch_softlockup_watchdog_sched(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400238{
Andrew Morton78611442014-04-18 15:07:12 -0700239 /*
240 * Preemption can be enabled. It doesn't matter which CPU's timestamp
241 * gets zeroed here, so use the raw_ operation.
242 */
243 raw_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400244}
Tejun Heo03e0d462015-12-08 11:28:04 -0500245
246void touch_softlockup_watchdog(void)
247{
248 touch_softlockup_watchdog_sched();
Tejun Heo82607adc2015-12-08 11:28:04 -0500249 wq_watchdog_touch(raw_smp_processor_id());
Tejun Heo03e0d462015-12-08 11:28:04 -0500250}
Ingo Molnar0167c782010-05-13 08:53:33 +0200251EXPORT_SYMBOL(touch_softlockup_watchdog);
Don Zickus58687ac2010-05-07 17:11:44 -0400252
Don Zickus332fbdb2010-05-07 17:11:45 -0400253void touch_all_softlockup_watchdogs(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400254{
255 int cpu;
256
257 /*
258 * this is done lockless
259 * do we care if a 0 races with a timestamp?
260 * all it means is the softlock check starts one cycle later
261 */
Chris Metcalffe4ba3c2015-06-24 16:55:45 -0700262 for_each_watchdog_cpu(cpu)
Don Zickus58687ac2010-05-07 17:11:44 -0400263 per_cpu(watchdog_touch_ts, cpu) = 0;
Tejun Heo82607adc2015-12-08 11:28:04 -0500264 wq_watchdog_touch(-1);
Don Zickus58687ac2010-05-07 17:11:44 -0400265}
266
Don Zickuscafcd802010-05-14 11:11:21 -0400267#ifdef CONFIG_HARDLOCKUP_DETECTOR
Don Zickus58687ac2010-05-07 17:11:44 -0400268void touch_nmi_watchdog(void)
269{
Ben Zhang62572e22014-04-03 14:47:18 -0700270 /*
271 * Using __raw here because some code paths have
272 * preemption enabled. If preemption is enabled
273 * then interrupts should be enabled too, in which
274 * case we shouldn't have to worry about the watchdog
275 * going off.
276 */
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500277 raw_cpu_write(watchdog_nmi_touch, true);
Don Zickus332fbdb2010-05-07 17:11:45 -0400278 touch_softlockup_watchdog();
Don Zickus58687ac2010-05-07 17:11:44 -0400279}
280EXPORT_SYMBOL(touch_nmi_watchdog);
281
Don Zickuscafcd802010-05-14 11:11:21 -0400282#endif
283
Don Zickus58687ac2010-05-07 17:11:44 -0400284void touch_softlockup_watchdog_sync(void)
285{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500286 __this_cpu_write(softlockup_touch_sync, true);
287 __this_cpu_write(watchdog_touch_ts, 0);
Don Zickus58687ac2010-05-07 17:11:44 -0400288}
289
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200290#ifdef CONFIG_HARDLOCKUP_DETECTOR
Don Zickus58687ac2010-05-07 17:11:44 -0400291/* watchdog detector functions */
Yaowei Bai451637e2015-11-05 18:44:24 -0800292static bool is_hardlockup(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400293{
Christoph Lameter909ea962010-12-08 16:22:55 +0100294 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400295
Christoph Lameter909ea962010-12-08 16:22:55 +0100296 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
Yaowei Bai451637e2015-11-05 18:44:24 -0800297 return true;
Don Zickus58687ac2010-05-07 17:11:44 -0400298
Christoph Lameter909ea962010-12-08 16:22:55 +0100299 __this_cpu_write(hrtimer_interrupts_saved, hrint);
Yaowei Bai451637e2015-11-05 18:44:24 -0800300 return false;
Don Zickus58687ac2010-05-07 17:11:44 -0400301}
302#endif
303
Don Zickus26e09c62010-05-17 18:06:04 -0400304static int is_softlockup(unsigned long touch_ts)
Don Zickus58687ac2010-05-07 17:11:44 -0400305{
Namhyung Kimc06b4f12012-12-27 11:49:44 +0900306 unsigned long now = get_timestamp();
Don Zickus58687ac2010-05-07 17:11:44 -0400307
Ulrich Obergfell39d2da22015-11-05 18:44:56 -0800308 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700309 /* Warn about unreasonable delays. */
310 if (time_after(now, touch_ts + get_softlockup_thresh()))
311 return now - touch_ts;
312 }
Don Zickus58687ac2010-05-07 17:11:44 -0400313 return 0;
314}
315
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200316#ifdef CONFIG_HARDLOCKUP_DETECTOR
Cyrill Gorcunov1880c4a2011-06-23 16:49:18 +0400317
Don Zickus58687ac2010-05-07 17:11:44 -0400318static struct perf_event_attr wd_hw_attr = {
319 .type = PERF_TYPE_HARDWARE,
320 .config = PERF_COUNT_HW_CPU_CYCLES,
321 .size = sizeof(struct perf_event_attr),
322 .pinned = 1,
323 .disabled = 1,
324};
325
326/* Callback function for perf event subsystem */
Peter Zijlstraa8b0ca12011-06-27 14:41:57 +0200327static void watchdog_overflow_callback(struct perf_event *event,
Don Zickus58687ac2010-05-07 17:11:44 -0400328 struct perf_sample_data *data,
329 struct pt_regs *regs)
330{
Peter Zijlstrac6db67c2010-08-20 11:49:15 +0200331 /* Ensure the watchdog never gets throttled */
332 event->hw.interrupts = 0;
333
Christoph Lameter909ea962010-12-08 16:22:55 +0100334 if (__this_cpu_read(watchdog_nmi_touch) == true) {
335 __this_cpu_write(watchdog_nmi_touch, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400336 return;
337 }
338
339 /* check for a hardlockup
340 * This is done by making sure our timer interrupt
341 * is incrementing. The timer interrupt should have
342 * fired multiple times before we overflow'd. If it hasn't
343 * then this is a good indication the cpu is stuck
344 */
Don Zickus26e09c62010-05-17 18:06:04 -0400345 if (is_hardlockup()) {
346 int this_cpu = smp_processor_id();
Jiri Kosina55537872015-11-05 18:44:41 -0800347 struct pt_regs *regs = get_irq_regs();
Don Zickus26e09c62010-05-17 18:06:04 -0400348
Don Zickus58687ac2010-05-07 17:11:44 -0400349 /* only print hardlockups once */
Christoph Lameter909ea962010-12-08 16:22:55 +0100350 if (__this_cpu_read(hard_watchdog_warn) == true)
Don Zickus58687ac2010-05-07 17:11:44 -0400351 return;
352
Jiri Kosina55537872015-11-05 18:44:41 -0800353 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
354 print_modules();
355 print_irqtrace_events(current);
356 if (regs)
357 show_regs(regs);
Don Zickus58687ac2010-05-07 17:11:44 -0400358 else
Jiri Kosina55537872015-11-05 18:44:41 -0800359 dump_stack();
360
361 /*
362 * Perform all-CPU dump only once to avoid multiple hardlockups
363 * generating interleaving traces
364 */
365 if (sysctl_hardlockup_all_cpu_backtrace &&
366 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
367 trigger_allbutself_cpu_backtrace();
368
369 if (hardlockup_panic)
Hidehiro Kawai58c56612015-12-14 11:19:10 +0100370 nmi_panic(regs, "Hard LOCKUP");
Don Zickus58687ac2010-05-07 17:11:44 -0400371
Christoph Lameter909ea962010-12-08 16:22:55 +0100372 __this_cpu_write(hard_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400373 return;
374 }
375
Christoph Lameter909ea962010-12-08 16:22:55 +0100376 __this_cpu_write(hard_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400377 return;
378}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000379#endif /* CONFIG_HARDLOCKUP_DETECTOR */
380
Don Zickus58687ac2010-05-07 17:11:44 -0400381static void watchdog_interrupt_count(void)
382{
Christoph Lameter909ea962010-12-08 16:22:55 +0100383 __this_cpu_inc(hrtimer_interrupts);
Don Zickus58687ac2010-05-07 17:11:44 -0400384}
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000385
386static int watchdog_nmi_enable(unsigned int cpu);
387static void watchdog_nmi_disable(unsigned int cpu);
Don Zickus58687ac2010-05-07 17:11:44 -0400388
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800389static int watchdog_enable_all_cpus(void);
390static void watchdog_disable_all_cpus(void);
391
Don Zickus58687ac2010-05-07 17:11:44 -0400392/* watchdog kicker functions */
393static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
394{
Christoph Lameter909ea962010-12-08 16:22:55 +0100395 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400396 struct pt_regs *regs = get_irq_regs();
397 int duration;
Aaron Tomlined235872014-06-23 13:22:05 -0700398 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
Don Zickus58687ac2010-05-07 17:11:44 -0400399
400 /* kick the hardlockup detector */
401 watchdog_interrupt_count();
402
403 /* kick the softlockup detector */
Christoph Lameter909ea962010-12-08 16:22:55 +0100404 wake_up_process(__this_cpu_read(softlockup_watchdog));
Don Zickus58687ac2010-05-07 17:11:44 -0400405
406 /* .. and repeat */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800407 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
Don Zickus58687ac2010-05-07 17:11:44 -0400408
409 if (touch_ts == 0) {
Christoph Lameter909ea962010-12-08 16:22:55 +0100410 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
Don Zickus58687ac2010-05-07 17:11:44 -0400411 /*
412 * If the time stamp was touched atomically
413 * make sure the scheduler tick is up to date.
414 */
Christoph Lameter909ea962010-12-08 16:22:55 +0100415 __this_cpu_write(softlockup_touch_sync, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400416 sched_clock_tick();
417 }
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500418
419 /* Clear the guest paused flag on watchdog reset */
420 kvm_check_and_clear_guest_paused();
Don Zickus58687ac2010-05-07 17:11:44 -0400421 __touch_watchdog();
422 return HRTIMER_RESTART;
423 }
424
425 /* check for a softlockup
426 * This is done by making sure a high priority task is
427 * being scheduled. The task touches the watchdog to
428 * indicate it is getting cpu time. If it hasn't then
429 * this is a good indication some task is hogging the cpu
430 */
Don Zickus26e09c62010-05-17 18:06:04 -0400431 duration = is_softlockup(touch_ts);
Don Zickus58687ac2010-05-07 17:11:44 -0400432 if (unlikely(duration)) {
Eric B Munson5d1c0f42012-03-10 14:37:28 -0500433 /*
434 * If a virtual machine is stopped by the host it can look to
435 * the watchdog like a soft lockup, check to see if the host
436 * stopped the vm before we issue the warning
437 */
438 if (kvm_check_and_clear_guest_paused())
439 return HRTIMER_RESTART;
440
Don Zickus58687ac2010-05-07 17:11:44 -0400441 /* only warn once */
chai wenb1a8de12014-10-09 15:25:17 -0700442 if (__this_cpu_read(soft_watchdog_warn) == true) {
443 /*
444 * When multiple processes are causing softlockups the
445 * softlockup detector only warns on the first one
446 * because the code relies on a full quiet cycle to
447 * re-arm. The second process prevents the quiet cycle
448 * and never gets reported. Use task pointers to detect
449 * this.
450 */
451 if (__this_cpu_read(softlockup_task_ptr_saved) !=
452 current) {
453 __this_cpu_write(soft_watchdog_warn, false);
454 __touch_watchdog();
455 }
Don Zickus58687ac2010-05-07 17:11:44 -0400456 return HRTIMER_RESTART;
chai wenb1a8de12014-10-09 15:25:17 -0700457 }
Don Zickus58687ac2010-05-07 17:11:44 -0400458
Aaron Tomlined235872014-06-23 13:22:05 -0700459 if (softlockup_all_cpu_backtrace) {
460 /* Prevent multiple soft-lockup reports if one cpu is already
461 * engaged in dumping cpu back traces
462 */
463 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
464 /* Someone else will report us. Let's give up */
465 __this_cpu_write(soft_watchdog_warn, true);
466 return HRTIMER_RESTART;
467 }
468 }
469
Fabian Frederick656c3b72014-08-06 16:04:03 -0700470 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
Don Zickus26e09c62010-05-17 18:06:04 -0400471 smp_processor_id(), duration,
Don Zickus58687ac2010-05-07 17:11:44 -0400472 current->comm, task_pid_nr(current));
chai wenb1a8de12014-10-09 15:25:17 -0700473 __this_cpu_write(softlockup_task_ptr_saved, current);
Don Zickus58687ac2010-05-07 17:11:44 -0400474 print_modules();
475 print_irqtrace_events(current);
476 if (regs)
477 show_regs(regs);
478 else
479 dump_stack();
480
Aaron Tomlined235872014-06-23 13:22:05 -0700481 if (softlockup_all_cpu_backtrace) {
482 /* Avoid generating two back traces for current
483 * given that one is already made above
484 */
485 trigger_allbutself_cpu_backtrace();
486
487 clear_bit(0, &soft_lockup_nmi_warn);
488 /* Barrier to sync with other cpus */
489 smp_mb__after_atomic();
490 }
491
Josh Hunt69361ee2014-08-08 14:22:31 -0700492 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
Don Zickus58687ac2010-05-07 17:11:44 -0400493 if (softlockup_panic)
494 panic("softlockup: hung tasks");
Christoph Lameter909ea962010-12-08 16:22:55 +0100495 __this_cpu_write(soft_watchdog_warn, true);
Don Zickus58687ac2010-05-07 17:11:44 -0400496 } else
Christoph Lameter909ea962010-12-08 16:22:55 +0100497 __this_cpu_write(soft_watchdog_warn, false);
Don Zickus58687ac2010-05-07 17:11:44 -0400498
499 return HRTIMER_RESTART;
500}
501
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000502static void watchdog_set_prio(unsigned int policy, unsigned int prio)
Don Zickus58687ac2010-05-07 17:11:44 -0400503{
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000504 struct sched_param param = { .sched_priority = prio };
505
506 sched_setscheduler(current, policy, &param);
507}
508
509static void watchdog_enable(unsigned int cpu)
510{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500511 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
Don Zickus58687ac2010-05-07 17:11:44 -0400512
Bjørn Mork3935e8952012-12-19 20:51:31 +0100513 /* kick off the timer for the hardlockup detector */
514 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
515 hrtimer->function = watchdog_timer_fn;
516
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000517 /* Enable the perf event */
518 watchdog_nmi_enable(cpu);
Don Zickus58687ac2010-05-07 17:11:44 -0400519
Don Zickus58687ac2010-05-07 17:11:44 -0400520 /* done here because hrtimer_start can only pin to smp_processor_id() */
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800521 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
Don Zickus58687ac2010-05-07 17:11:44 -0400522 HRTIMER_MODE_REL_PINNED);
523
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000524 /* initialize timestamp */
525 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
526 __touch_watchdog();
Don Zickus58687ac2010-05-07 17:11:44 -0400527}
528
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000529static void watchdog_disable(unsigned int cpu)
530{
Christoph Lameterf7f66b02014-08-17 12:30:34 -0500531 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000532
533 watchdog_set_prio(SCHED_NORMAL, 0);
534 hrtimer_cancel(hrtimer);
535 /* disable the perf event */
536 watchdog_nmi_disable(cpu);
537}
538
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200539static void watchdog_cleanup(unsigned int cpu, bool online)
540{
541 watchdog_disable(cpu);
542}
543
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000544static int watchdog_should_run(unsigned int cpu)
545{
546 return __this_cpu_read(hrtimer_interrupts) !=
547 __this_cpu_read(soft_lockup_hrtimer_cnt);
548}
549
550/*
551 * The watchdog thread function - touches the timestamp.
552 *
Chuansheng Liu0f34c402012-12-17 15:59:50 -0800553 * It only runs once every sample_period seconds (4 seconds by
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000554 * default) to reset the softlockup timestamp. If this gets delayed
555 * for more than 2*watchdog_thresh seconds then the debug-printout
556 * triggers in watchdog_timer_fn().
557 */
558static void watchdog(unsigned int cpu)
559{
560 __this_cpu_write(soft_lockup_hrtimer_cnt,
561 __this_cpu_read(hrtimer_interrupts));
562 __touch_watchdog();
Ulrich Obergfellbcfba4f2015-04-14 15:44:10 -0700563
564 /*
565 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
566 * failure path. Check for failures that can occur asynchronously -
567 * for example, when CPUs are on-lined - and shut down the hardware
568 * perf event on each CPU accordingly.
569 *
570 * The only non-obvious place this bit can be cleared is through
571 * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a
572 * pr_info here would be too noisy as it would result in a message
573 * every few seconds if the hardlockup was disabled but the softlockup
574 * enabled.
575 */
576 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
577 watchdog_nmi_disable(cpu);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000578}
Don Zickus58687ac2010-05-07 17:11:44 -0400579
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200580#ifdef CONFIG_HARDLOCKUP_DETECTOR
Don Zickusa7027042012-06-13 09:35:48 -0400581/*
582 * People like the simple clean cpu node info on boot.
583 * Reduce the watchdog noise by only printing messages
584 * that are different from what cpu0 displayed.
585 */
586static unsigned long cpu0_err;
587
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000588static int watchdog_nmi_enable(unsigned int cpu)
Don Zickus58687ac2010-05-07 17:11:44 -0400589{
590 struct perf_event_attr *wd_attr;
591 struct perf_event *event = per_cpu(watchdog_ev, cpu);
592
Ulrich Obergfell195daf62015-04-14 15:44:13 -0700593 /* nothing to do if the hard lockup detector is disabled */
594 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
595 goto out;
Ulrich Obergfell6e7458a2014-10-13 15:55:35 -0700596
Don Zickus58687ac2010-05-07 17:11:44 -0400597 /* is it already setup and enabled? */
598 if (event && event->state > PERF_EVENT_STATE_OFF)
599 goto out;
600
601 /* it is setup but not enabled */
602 if (event != NULL)
603 goto out_enable;
604
Don Zickus58687ac2010-05-07 17:11:44 -0400605 wd_attr = &wd_hw_attr;
Mandeep Singh Baines4eec42f2011-05-22 22:10:23 -0700606 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
Cyrill Gorcunov1880c4a2011-06-23 16:49:18 +0400607
608 /* Try to register using hardware perf events */
Avi Kivity4dc0da82011-06-29 18:42:35 +0300609 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
Don Zickusa7027042012-06-13 09:35:48 -0400610
611 /* save cpu0 error for future comparision */
612 if (cpu == 0 && IS_ERR(event))
613 cpu0_err = PTR_ERR(event);
614
Don Zickus58687ac2010-05-07 17:11:44 -0400615 if (!IS_ERR(event)) {
Don Zickusa7027042012-06-13 09:35:48 -0400616 /* only print for cpu0 or different than cpu0 */
617 if (cpu == 0 || cpu0_err)
618 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
Don Zickus58687ac2010-05-07 17:11:44 -0400619 goto out_save;
620 }
621
Ulrich Obergfellbcfba4f2015-04-14 15:44:10 -0700622 /*
623 * Disable the hard lockup detector if _any_ CPU fails to set up
624 * set up the hardware perf event. The watchdog() function checks
625 * the NMI_WATCHDOG_ENABLED bit periodically.
626 *
627 * The barriers are for syncing up watchdog_enabled across all the
628 * cpus, as clear_bit() does not use barriers.
629 */
630 smp_mb__before_atomic();
631 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
632 smp_mb__after_atomic();
633
Don Zickusa7027042012-06-13 09:35:48 -0400634 /* skip displaying the same error again */
635 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
636 return PTR_ERR(event);
Don Zickus5651f7f2011-02-09 14:02:33 -0500637
638 /* vary the KERN level based on the returned errno */
639 if (PTR_ERR(event) == -EOPNOTSUPP)
Andrew Morton45019802012-03-23 15:01:55 -0700640 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
Don Zickus5651f7f2011-02-09 14:02:33 -0500641 else if (PTR_ERR(event) == -ENOENT)
Fabian Frederick656c3b72014-08-06 16:04:03 -0700642 pr_warn("disabled (cpu%i): hardware events not enabled\n",
Andrew Morton45019802012-03-23 15:01:55 -0700643 cpu);
Don Zickus5651f7f2011-02-09 14:02:33 -0500644 else
Andrew Morton45019802012-03-23 15:01:55 -0700645 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
646 cpu, PTR_ERR(event));
Ulrich Obergfellbcfba4f2015-04-14 15:44:10 -0700647
648 pr_info("Shutting down hard lockup detector on all cpus\n");
649
Akinobu Mitaeac24332010-08-31 23:00:08 -0400650 return PTR_ERR(event);
Don Zickus58687ac2010-05-07 17:11:44 -0400651
652 /* success path */
653out_save:
654 per_cpu(watchdog_ev, cpu) = event;
655out_enable:
656 perf_event_enable(per_cpu(watchdog_ev, cpu));
657out:
658 return 0;
659}
660
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000661static void watchdog_nmi_disable(unsigned int cpu)
Don Zickus58687ac2010-05-07 17:11:44 -0400662{
663 struct perf_event *event = per_cpu(watchdog_ev, cpu);
664
665 if (event) {
666 perf_event_disable(event);
667 per_cpu(watchdog_ev, cpu) = NULL;
668
669 /* should be in cleanup, but blocks oprofile */
670 perf_event_release_kernel(event);
671 }
Ulrich Obergfelldf577142014-08-11 10:49:25 -0400672 if (cpu == 0) {
673 /* watchdog_nmi_enable() expects this to be zero initially. */
674 cpu0_err = 0;
675 }
Don Zickus58687ac2010-05-07 17:11:44 -0400676}
Stephane Eranianb3738d22014-11-17 20:07:03 +0100677
Don Zickus58687ac2010-05-07 17:11:44 -0400678#else
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000679static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
680static void watchdog_nmi_disable(unsigned int cpu) { return; }
Frederic Weisbecker23637d42010-05-15 23:15:20 +0200681#endif /* CONFIG_HARDLOCKUP_DETECTOR */
Don Zickus58687ac2010-05-07 17:11:44 -0400682
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200683static struct smp_hotplug_thread watchdog_threads = {
684 .store = &softlockup_watchdog,
685 .thread_should_run = watchdog_should_run,
686 .thread_fn = watchdog,
687 .thread_comm = "watchdog/%u",
688 .setup = watchdog_enable,
689 .cleanup = watchdog_cleanup,
690 .park = watchdog_disable,
691 .unpark = watchdog_enable,
692};
693
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700694/*
695 * park all watchdog threads that are specified in 'watchdog_cpumask'
Ulrich Obergfellee7fed52015-11-05 18:44:39 -0800696 *
697 * This function returns an error if kthread_park() of a watchdog thread
698 * fails. In this situation, the watchdog threads of some CPUs can already
699 * be parked and the watchdog threads of other CPUs can still be runnable.
700 * Callers are expected to handle this special condition as appropriate in
701 * their context.
Ulrich Obergfella2a45b82015-11-05 18:44:53 -0800702 *
703 * This function may only be called in a context that is protected against
704 * races with CPU hotplug - for example, via get_online_cpus().
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700705 */
706static int watchdog_park_threads(void)
707{
708 int cpu, ret = 0;
709
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700710 for_each_watchdog_cpu(cpu) {
711 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
712 if (ret)
713 break;
714 }
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700715
716 return ret;
717}
718
719/*
720 * unpark all watchdog threads that are specified in 'watchdog_cpumask'
Ulrich Obergfella2a45b82015-11-05 18:44:53 -0800721 *
722 * This function may only be called in a context that is protected against
723 * races with CPU hotplug - for example, via get_online_cpus().
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700724 */
725static void watchdog_unpark_threads(void)
726{
727 int cpu;
728
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700729 for_each_watchdog_cpu(cpu)
730 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
Ulrich Obergfell81a4bee2015-09-04 15:45:15 -0700731}
732
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700733/*
734 * Suspend the hard and soft lockup detector by parking the watchdog threads.
735 */
Ulrich Obergfellec6a9062015-09-04 15:45:28 -0700736int lockup_detector_suspend(void)
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700737{
738 int ret = 0;
739
Ulrich Obergfellee89e712015-11-05 18:44:47 -0800740 get_online_cpus();
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700741 mutex_lock(&watchdog_proc_mutex);
742 /*
743 * Multiple suspend requests can be active in parallel (counted by
744 * the 'watchdog_suspended' variable). If the watchdog threads are
745 * running, the first caller takes care that they will be parked.
746 * The state of 'watchdog_running' cannot change while a suspend
Ulrich Obergfellec6a9062015-09-04 15:45:28 -0700747 * request is active (see related code in 'proc' handlers).
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700748 */
749 if (watchdog_running && !watchdog_suspended)
750 ret = watchdog_park_threads();
751
752 if (ret == 0)
753 watchdog_suspended++;
Ulrich Obergfellc9935902015-11-05 18:44:36 -0800754 else {
755 watchdog_disable_all_cpus();
756 pr_err("Failed to suspend lockup detectors, disabled\n");
757 watchdog_enabled = 0;
758 }
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700759
760 mutex_unlock(&watchdog_proc_mutex);
761
762 return ret;
763}
764
765/*
766 * Resume the hard and soft lockup detector by unparking the watchdog threads.
767 */
Ulrich Obergfellec6a9062015-09-04 15:45:28 -0700768void lockup_detector_resume(void)
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700769{
770 mutex_lock(&watchdog_proc_mutex);
771
772 watchdog_suspended--;
773 /*
774 * The watchdog threads are unparked if they were previously running
775 * and if there is no more active suspend request.
776 */
777 if (watchdog_running && !watchdog_suspended)
778 watchdog_unpark_threads();
779
780 mutex_unlock(&watchdog_proc_mutex);
Ulrich Obergfellee89e712015-11-05 18:44:47 -0800781 put_online_cpus();
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700782}
783
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800784static int update_watchdog_all_cpus(void)
Michal Hocko9809b182013-09-24 15:27:30 -0700785{
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800786 int ret;
787
788 ret = watchdog_park_threads();
789 if (ret)
790 return ret;
791
Ulrich Obergfelld4bdd0b2015-09-04 15:45:21 -0700792 watchdog_unpark_threads();
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800793
794 return 0;
Michal Hocko9809b182013-09-24 15:27:30 -0700795}
796
Ulrich Obergfellb2f57c32015-04-14 15:44:16 -0700797static int watchdog_enable_all_cpus(void)
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200798{
799 int err = 0;
800
Frederic Weisbecker3c00ea82013-05-19 20:45:15 +0200801 if (!watchdog_running) {
Frederic Weisbecker230ec932015-09-04 15:45:06 -0700802 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
803 &watchdog_cpumask);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200804 if (err)
805 pr_err("Failed to create watchdog threads, disabled\n");
Frederic Weisbecker230ec932015-09-04 15:45:06 -0700806 else
Frederic Weisbecker3c00ea82013-05-19 20:45:15 +0200807 watchdog_running = 1;
Ulrich Obergfellb2f57c32015-04-14 15:44:16 -0700808 } else {
809 /*
810 * Enable/disable the lockup detectors or
811 * change the sample period 'on the fly'.
812 */
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800813 err = update_watchdog_all_cpus();
814
815 if (err) {
816 watchdog_disable_all_cpus();
817 pr_err("Failed to update lockup detectors, disabled\n");
818 }
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200819 }
820
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800821 if (err)
822 watchdog_enabled = 0;
823
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200824 return err;
825}
826
Don Zickus58687ac2010-05-07 17:11:44 -0400827static void watchdog_disable_all_cpus(void)
828{
Frederic Weisbecker3c00ea82013-05-19 20:45:15 +0200829 if (watchdog_running) {
830 watchdog_running = 0;
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200831 smpboot_unregister_percpu_thread(&watchdog_threads);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000832 }
Don Zickus58687ac2010-05-07 17:11:44 -0400833}
834
Ulrich Obergfell58cf6902015-11-05 18:44:30 -0800835#ifdef CONFIG_SYSCTL
836
Don Zickus58687ac2010-05-07 17:11:44 -0400837/*
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700838 * Update the run state of the lockup detectors.
Don Zickus58687ac2010-05-07 17:11:44 -0400839 */
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700840static int proc_watchdog_update(void)
Don Zickus58687ac2010-05-07 17:11:44 -0400841{
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700842 int err = 0;
843
844 /*
845 * Watchdog threads won't be started if they are already active.
846 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
847 * care of this. If those threads are already active, the sample
848 * period will be updated and the lockup detectors will be enabled
849 * or disabled 'on the fly'.
850 */
851 if (watchdog_enabled && watchdog_thresh)
Ulrich Obergfellb2f57c32015-04-14 15:44:16 -0700852 err = watchdog_enable_all_cpus();
Ulrich Obergfella0c9cbb2015-04-14 15:43:58 -0700853 else
854 watchdog_disable_all_cpus();
855
856 return err;
857
858}
859
860/*
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700861 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
862 *
863 * caller | table->data points to | 'which' contains the flag(s)
864 * -------------------|-----------------------|-----------------------------
865 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
866 * | | with SOFT_WATCHDOG_ENABLED
867 * -------------------|-----------------------|-----------------------------
868 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
869 * -------------------|-----------------------|-----------------------------
870 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
871 */
872static int proc_watchdog_common(int which, struct ctl_table *table, int write,
873 void __user *buffer, size_t *lenp, loff_t *ppos)
874{
875 int err, old, new;
876 int *watchdog_param = (int *)table->data;
Don Zickus58687ac2010-05-07 17:11:44 -0400877
Ulrich Obergfell8614dde2015-11-05 18:44:50 -0800878 get_online_cpus();
Michal Hocko359e6fa2013-09-24 15:27:29 -0700879 mutex_lock(&watchdog_proc_mutex);
Thomas Gleixnerbcd951c2012-07-16 10:42:38 +0000880
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700881 if (watchdog_suspended) {
882 /* no parameter changes allowed while watchdog is suspended */
883 err = -EAGAIN;
884 goto out;
885 }
886
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700887 /*
888 * If the parameter is being read return the state of the corresponding
889 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
890 * run state of the lockup detectors.
891 */
892 if (!write) {
893 *watchdog_param = (watchdog_enabled & which) != 0;
894 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
895 } else {
896 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
897 if (err)
898 goto out;
899
900 /*
901 * There is a race window between fetching the current value
902 * from 'watchdog_enabled' and storing the new value. During
903 * this race window, watchdog_nmi_enable() can sneak in and
904 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
905 * The 'cmpxchg' detects this race and the loop retries.
906 */
907 do {
908 old = watchdog_enabled;
909 /*
910 * If the parameter value is not zero set the
911 * corresponding bit(s), else clear it(them).
912 */
913 if (*watchdog_param)
914 new = old | which;
915 else
916 new = old & ~which;
917 } while (cmpxchg(&watchdog_enabled, old, new) != old);
918
919 /*
Ulrich Obergfellb43cb432015-11-05 18:44:33 -0800920 * Update the run state of the lockup detectors. There is _no_
921 * need to check the value returned by proc_watchdog_update()
922 * and to restore the previous value of 'watchdog_enabled' as
923 * both lockup detectors are disabled if proc_watchdog_update()
924 * returns an error.
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700925 */
Joshua Hunta1ee1932016-03-17 14:17:23 -0700926 if (old == new)
927 goto out;
928
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700929 err = proc_watchdog_update();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700930 }
931out:
932 mutex_unlock(&watchdog_proc_mutex);
Ulrich Obergfell8614dde2015-11-05 18:44:50 -0800933 put_online_cpus();
Ulrich Obergfellef246a22015-04-14 15:44:05 -0700934 return err;
935}
936
937/*
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700938 * /proc/sys/kernel/watchdog
939 */
940int proc_watchdog(struct ctl_table *table, int write,
941 void __user *buffer, size_t *lenp, loff_t *ppos)
942{
943 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
944 table, write, buffer, lenp, ppos);
945}
946
947/*
948 * /proc/sys/kernel/nmi_watchdog
949 */
950int proc_nmi_watchdog(struct ctl_table *table, int write,
951 void __user *buffer, size_t *lenp, loff_t *ppos)
952{
953 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
954 table, write, buffer, lenp, ppos);
955}
956
957/*
958 * /proc/sys/kernel/soft_watchdog
959 */
960int proc_soft_watchdog(struct ctl_table *table, int write,
961 void __user *buffer, size_t *lenp, loff_t *ppos)
962{
963 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
964 table, write, buffer, lenp, ppos);
965}
966
967/*
968 * /proc/sys/kernel/watchdog_thresh
969 */
970int proc_watchdog_thresh(struct ctl_table *table, int write,
971 void __user *buffer, size_t *lenp, loff_t *ppos)
972{
Joshua Hunta1ee1932016-03-17 14:17:23 -0700973 int err, old, new;
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700974
Ulrich Obergfell8614dde2015-11-05 18:44:50 -0800975 get_online_cpus();
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700976 mutex_lock(&watchdog_proc_mutex);
977
Ulrich Obergfell8c073d22015-09-04 15:45:18 -0700978 if (watchdog_suspended) {
979 /* no parameter changes allowed while watchdog is suspended */
980 err = -EAGAIN;
981 goto out;
982 }
983
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700984 old = ACCESS_ONCE(watchdog_thresh);
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200985 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700986
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +0200987 if (err || !write)
Michal Hocko359e6fa2013-09-24 15:27:29 -0700988 goto out;
Mandeep Singh Bainese04ab2b2011-05-22 22:10:21 -0700989
anish kumarb66a2352013-03-12 14:44:08 -0400990 /*
Ulrich Obergfelld283c642015-11-05 18:44:27 -0800991 * Update the sample period. Restore on failure.
anish kumarb66a2352013-03-12 14:44:08 -0400992 */
Joshua Hunta1ee1932016-03-17 14:17:23 -0700993 new = ACCESS_ONCE(watchdog_thresh);
994 if (old == new)
995 goto out;
996
Ulrich Obergfell83a80a32015-04-14 15:44:08 -0700997 set_sample_period();
998 err = proc_watchdog_update();
Ulrich Obergfelld283c642015-11-05 18:44:27 -0800999 if (err) {
Ulrich Obergfell83a80a32015-04-14 15:44:08 -07001000 watchdog_thresh = old;
Ulrich Obergfelld283c642015-11-05 18:44:27 -08001001 set_sample_period();
1002 }
Michal Hocko359e6fa2013-09-24 15:27:29 -07001003out:
1004 mutex_unlock(&watchdog_proc_mutex);
Ulrich Obergfell8614dde2015-11-05 18:44:50 -08001005 put_online_cpus();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +02001006 return err;
Don Zickus58687ac2010-05-07 17:11:44 -04001007}
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001008
1009/*
1010 * The cpumask is the mask of possible cpus that the watchdog can run
1011 * on, not the mask of cpus it is actually running on. This allows the
1012 * user to specify a mask that will include cpus that have not yet
1013 * been brought online, if desired.
1014 */
1015int proc_watchdog_cpumask(struct ctl_table *table, int write,
1016 void __user *buffer, size_t *lenp, loff_t *ppos)
1017{
1018 int err;
1019
Ulrich Obergfell8614dde2015-11-05 18:44:50 -08001020 get_online_cpus();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001021 mutex_lock(&watchdog_proc_mutex);
Ulrich Obergfell8c073d22015-09-04 15:45:18 -07001022
1023 if (watchdog_suspended) {
1024 /* no parameter changes allowed while watchdog is suspended */
1025 err = -EAGAIN;
1026 goto out;
1027 }
1028
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001029 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1030 if (!err && write) {
1031 /* Remove impossible cpus to keep sysctl output cleaner. */
1032 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1033 cpu_possible_mask);
1034
1035 if (watchdog_running) {
1036 /*
1037 * Failure would be due to being unable to allocate
1038 * a temporary cpumask, so we are likely not in a
1039 * position to do much else to make things better.
1040 */
1041 if (smpboot_update_cpumask_percpu_thread(
1042 &watchdog_threads, &watchdog_cpumask) != 0)
1043 pr_err("cpumask update failed\n");
1044 }
1045 }
Ulrich Obergfell8c073d22015-09-04 15:45:18 -07001046out:
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001047 mutex_unlock(&watchdog_proc_mutex);
Ulrich Obergfell8614dde2015-11-05 18:44:50 -08001048 put_online_cpus();
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001049 return err;
1050}
1051
Don Zickus58687ac2010-05-07 17:11:44 -04001052#endif /* CONFIG_SYSCTL */
1053
Peter Zijlstra004417a2010-11-25 18:38:29 +01001054void __init lockup_detector_init(void)
Don Zickus58687ac2010-05-07 17:11:44 -04001055{
Chuansheng Liu0f34c402012-12-17 15:59:50 -08001056 set_sample_period();
Frederic Weisbeckerb8900bc2013-06-06 15:42:53 +02001057
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001058#ifdef CONFIG_NO_HZ_FULL
1059 if (tick_nohz_full_enabled()) {
Frederic Weisbecker314b08ff2015-09-04 15:45:09 -07001060 pr_info("Disabling watchdog on nohz_full cores by default\n");
1061 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
Chris Metcalffe4ba3c2015-06-24 16:55:45 -07001062 } else
1063 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1064#else
1065 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1066#endif
1067
Ulrich Obergfell195daf62015-04-14 15:44:13 -07001068 if (watchdog_enabled)
Ulrich Obergfellb2f57c32015-04-14 15:44:16 -07001069 watchdog_enable_all_cpus();
Don Zickus58687ac2010-05-07 17:11:44 -04001070}