linux/kernel/watchdog.c
<<
>>
Prefs
   1/*
   2 * Detect hard and soft lockups on a system
   3 *
   4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   5 *
   6 * 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
   9 * to those contributors as well.
  10 */
  11
  12#define pr_fmt(fmt) "NMI watchdog: " fmt
  13
  14#include <linux/mm.h>
  15#include <linux/cpu.h>
  16#include <linux/nmi.h>
  17#include <linux/init.h>
  18#include <linux/delay.h>
  19#include <linux/freezer.h>
  20#include <linux/kthread.h>
  21#include <linux/lockdep.h>
  22#include <linux/notifier.h>
  23#include <linux/module.h>
  24#include <linux/sysctl.h>
  25#include <linux/smpboot.h>
  26
  27#include <asm/irq_regs.h>
  28#include <linux/kvm_para.h>
  29#include <linux/perf_event.h>
  30
  31int watchdog_enabled = 1;
  32int __read_mostly watchdog_thresh = 10;
  33static int __read_mostly watchdog_disabled;
  34
  35static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
  36static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
  37static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
  38static DEFINE_PER_CPU(bool, softlockup_touch_sync);
  39static DEFINE_PER_CPU(bool, soft_watchdog_warn);
  40static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
  41static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
  42#ifdef CONFIG_HARDLOCKUP_DETECTOR
  43static DEFINE_PER_CPU(bool, hard_watchdog_warn);
  44static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
  45static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
  46static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
  47#endif
  48
  49/* boot commands */
  50/*
  51 * Should we panic when a soft-lockup or hard-lockup occurs:
  52 */
  53#ifdef CONFIG_HARDLOCKUP_DETECTOR
  54static int hardlockup_panic =
  55                        CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
  56
  57static int __init hardlockup_panic_setup(char *str)
  58{
  59        if (!strncmp(str, "panic", 5))
  60                hardlockup_panic = 1;
  61        else if (!strncmp(str, "nopanic", 7))
  62                hardlockup_panic = 0;
  63        else if (!strncmp(str, "0", 1))
  64                watchdog_enabled = 0;
  65        return 1;
  66}
  67__setup("nmi_watchdog=", hardlockup_panic_setup);
  68#endif
  69
  70unsigned int __read_mostly softlockup_panic =
  71                        CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  72
  73static int __init softlockup_panic_setup(char *str)
  74{
  75        softlockup_panic = simple_strtoul(str, NULL, 0);
  76
  77        return 1;
  78}
  79__setup("softlockup_panic=", softlockup_panic_setup);
  80
  81static int __init nowatchdog_setup(char *str)
  82{
  83        watchdog_enabled = 0;
  84        return 1;
  85}
  86__setup("nowatchdog", nowatchdog_setup);
  87
  88/* deprecated */
  89static int __init nosoftlockup_setup(char *str)
  90{
  91        watchdog_enabled = 0;
  92        return 1;
  93}
  94__setup("nosoftlockup", nosoftlockup_setup);
  95/*  */
  96
  97/*
  98 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
  99 * lockups can have false positives under extreme conditions. So we generally
 100 * want a higher threshold for soft lockups than for hard lockups. So we couple
 101 * the thresholds with a factor: we make the soft threshold twice the amount of
 102 * time the hard threshold is.
 103 */
 104static int get_softlockup_thresh(void)
 105{
 106        return watchdog_thresh * 2;
 107}
 108
 109/*
 110 * Returns seconds, approximately.  We don't need nanosecond
 111 * resolution, and we don't need to waste time with a big divide when
 112 * 2^30ns == 1.074s.
 113 */
 114static unsigned long get_timestamp(int this_cpu)
 115{
 116        return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
 117}
 118
 119static u64 get_sample_period(void)
 120{
 121        /*
 122         * convert watchdog_thresh from seconds to ns
 123         * the divide by 5 is to give hrtimer several chances (two
 124         * or three with the current relation between the soft
 125         * and hard thresholds) to increment before the
 126         * hardlockup detector generates a warning
 127         */
 128        return get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
 129}
 130
 131/* Commands for resetting the watchdog */
 132static void __touch_watchdog(void)
 133{
 134        int this_cpu = smp_processor_id();
 135
 136        __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
 137}
 138
 139void touch_softlockup_watchdog(void)
 140{
 141        __this_cpu_write(watchdog_touch_ts, 0);
 142}
 143EXPORT_SYMBOL(touch_softlockup_watchdog);
 144
 145void touch_all_softlockup_watchdogs(void)
 146{
 147        int cpu;
 148
 149        /*
 150         * this is done lockless
 151         * do we care if a 0 races with a timestamp?
 152         * all it means is the softlock check starts one cycle later
 153         */
 154        for_each_online_cpu(cpu)
 155                per_cpu(watchdog_touch_ts, cpu) = 0;
 156}
 157
 158#ifdef CONFIG_HARDLOCKUP_DETECTOR
 159void touch_nmi_watchdog(void)
 160{
 161        if (watchdog_enabled) {
 162                unsigned cpu;
 163
 164                for_each_present_cpu(cpu) {
 165                        if (per_cpu(watchdog_nmi_touch, cpu) != true)
 166                                per_cpu(watchdog_nmi_touch, cpu) = true;
 167                }
 168        }
 169        touch_softlockup_watchdog();
 170}
 171EXPORT_SYMBOL(touch_nmi_watchdog);
 172
 173#endif
 174
 175void touch_softlockup_watchdog_sync(void)
 176{
 177        __raw_get_cpu_var(softlockup_touch_sync) = true;
 178        __raw_get_cpu_var(watchdog_touch_ts) = 0;
 179}
 180
 181#ifdef CONFIG_HARDLOCKUP_DETECTOR
 182/* watchdog detector functions */
 183static int is_hardlockup(void)
 184{
 185        unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
 186
 187        if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
 188                return 1;
 189
 190        __this_cpu_write(hrtimer_interrupts_saved, hrint);
 191        return 0;
 192}
 193#endif
 194
 195static int is_softlockup(unsigned long touch_ts)
 196{
 197        unsigned long now = get_timestamp(smp_processor_id());
 198
 199        /* Warn about unreasonable delays: */
 200        if (time_after(now, touch_ts + get_softlockup_thresh()))
 201                return now - touch_ts;
 202
 203        return 0;
 204}
 205
 206#ifdef CONFIG_HARDLOCKUP_DETECTOR
 207
 208static struct perf_event_attr wd_hw_attr = {
 209        .type           = PERF_TYPE_HARDWARE,
 210        .config         = PERF_COUNT_HW_CPU_CYCLES,
 211        .size           = sizeof(struct perf_event_attr),
 212        .pinned         = 1,
 213        .disabled       = 1,
 214};
 215
 216/* Callback function for perf event subsystem */
 217static void watchdog_overflow_callback(struct perf_event *event,
 218                 struct perf_sample_data *data,
 219                 struct pt_regs *regs)
 220{
 221        /* Ensure the watchdog never gets throttled */
 222        event->hw.interrupts = 0;
 223
 224        if (__this_cpu_read(watchdog_nmi_touch) == true) {
 225                __this_cpu_write(watchdog_nmi_touch, false);
 226                return;
 227        }
 228
 229        /* check for a hardlockup
 230         * This is done by making sure our timer interrupt
 231         * is incrementing.  The timer interrupt should have
 232         * fired multiple times before we overflow'd.  If it hasn't
 233         * then this is a good indication the cpu is stuck
 234         */
 235        if (is_hardlockup()) {
 236                int this_cpu = smp_processor_id();
 237
 238                /* only print hardlockups once */
 239                if (__this_cpu_read(hard_watchdog_warn) == true)
 240                        return;
 241
 242                if (hardlockup_panic)
 243                        panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 244                else
 245                        WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
 246
 247                __this_cpu_write(hard_watchdog_warn, true);
 248                return;
 249        }
 250
 251        __this_cpu_write(hard_watchdog_warn, false);
 252        return;
 253}
 254#endif /* CONFIG_HARDLOCKUP_DETECTOR */
 255
 256static void watchdog_interrupt_count(void)
 257{
 258        __this_cpu_inc(hrtimer_interrupts);
 259}
 260
 261static int watchdog_nmi_enable(unsigned int cpu);
 262static void watchdog_nmi_disable(unsigned int cpu);
 263
 264/* watchdog kicker functions */
 265static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 266{
 267        unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
 268        struct pt_regs *regs = get_irq_regs();
 269        int duration;
 270
 271        /* kick the hardlockup detector */
 272        watchdog_interrupt_count();
 273
 274        /* kick the softlockup detector */
 275        wake_up_process(__this_cpu_read(softlockup_watchdog));
 276
 277        /* .. and repeat */
 278        hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
 279
 280        if (touch_ts == 0) {
 281                if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 282                        /*
 283                         * If the time stamp was touched atomically
 284                         * make sure the scheduler tick is up to date.
 285                         */
 286                        __this_cpu_write(softlockup_touch_sync, false);
 287                        sched_clock_tick();
 288                }
 289
 290                /* Clear the guest paused flag on watchdog reset */
 291                kvm_check_and_clear_guest_paused();
 292                __touch_watchdog();
 293                return HRTIMER_RESTART;
 294        }
 295
 296        /* check for a softlockup
 297         * This is done by making sure a high priority task is
 298         * being scheduled.  The task touches the watchdog to
 299         * indicate it is getting cpu time.  If it hasn't then
 300         * this is a good indication some task is hogging the cpu
 301         */
 302        duration = is_softlockup(touch_ts);
 303        if (unlikely(duration)) {
 294      7l/wa" id="L303" class="line" name="L303"> 30fame="L285"> 285        a virtualhischne"cationoppthe Thment"o Clcput"> looks the watchdog to
 126/* ch than f,s="commer se   * ment"o Che watchdog to
 297/*generates a warning
 298         */
 239                if (kvm_check_and_clear_guesf">hrtimer)
 240                        return HRTIMER_RESTART;
 241
 292                 293                if (__this_cpu_read(<* chef="+code=hard_watchdog_warn"* chef="+code=hardsref"hdog_warn) == true)
HRTIMER_RESTART;
 215
 286         struct (WARN(1,  215
 287                        smp_proctouch_sync,  215
 218                        event->< , ( 219         struct ();
 290         struct  281                if (regs)
 282                        ( 293                else
, ();
 215
 226                if (< class="srea href="+code=panic" class="srea hress="sref">regs)
 287                        panic(();
 238                __this_cpu_write(true);
 2}               else
 290         struct __this_cpu_write(false);
 241
 252        return HRTIMER_RESTART;
 253}
 194
 262static void (unsigned int 

regs) 266{ 268 struct < classparapref="+code=cpu" classparaprtimer .duration =

214}; 138 169 < classsetmake sureref="+code=cpu" classsetmake sure_this_cpu_read(event->false); 170} 241 262static void 241 266{ 2_timer_fn(struct hrtimer *d&>event->__raw_get_cpu_var(false); 215 296 /* kick the hardlockup detector */ 177 < class="srihrtimer_restart" class="srihaw_get_cpu_var(hrtimer, hrtimer, < class="MODE"sre=EXPORT_SYMBOL" class="MODE"srertimref">false); 278 event->duration = watchdoef">false); 289 280static void watchdog_enabled) { 291 ( 282 52 return; 2L170"> 170} 194 121 278 < class="sref">watchtchdog_enabled" class="sref">watchdog_n(unsigned int 157 251 hut"> ="comminmer sclass="sref">sm()fckup detector */ 169 < class="ef">hrtimer_restart" class="ef">hrtimorward_now(hrtimer, ns_to_ktime(get_sampledL215"> 215 240 >hrtimer, < class="MODE"sre_PINNED=EXPORT_SYMBOL" class="MODE"sre_PINNEDass=" return; 241 151 278 < class="sset_g sunterrupt_count" class="sset_g suthis_78 hrtimer, 1 struct __touch_watchdog(); 170} 276 217static void 276 ) { 2_timer_fn(struct hrtimer *d&>event->__raw_get_cpu_var(false/a>) { 270 251 < class="sset_g sunterrupt_count" class="sset_g suthis_78 false/a>) { 272 < class="t">celrtimer_restart" class="t">celrtimorward_now(false/a>) { 151 watchdtment function fckup detector */ 1 struct < class="sref"=watchdog_nmi_disable" class="sref">watchdog_nm(unsigned int 170} 276 261static int 276 ) { 252 return __this_cpu_read(hrtimer_i !=enabled) { 290 struct __this_cpu_read(<* chess="sre class="twiref="+code=cpu"* chess="sre class="twiass=" return; 170} 202 283 294 7l/w/watchdog3c#L284" id="L284" class=3line"3name="L284"> 284/* Call-led. The tasired If is up to date. 285 294 7l/w/watchdog3c#L286" id="L286" class=3line"38ame="L126"> 126get() seconds (4 seconds bched atomically 297efault)mer g on wment">/* kick tired If * indist unrewatcd atomically 2982* class="srh_sof seconds t"> ut ubug- ent"ouChe watchdog to 299 class="sref">wat() is up to date. 300 217static void 276 ) { 278 __this_cpu_write( 215 __this_cpu_read(hrtimer_i" return; 275 __touch_watchdog(); 170} 157 206#ifdef CONFIG_HARDLOCKUP_DETECTOR 299 294 7l/4/watchdog4c#L300" id="L300" class=4line"4name="L300"> 300 301/* Ensure ois * Th="comment"e taed sageriority task is 232wapnrewa is up to date. 283 67 unsigned long ref="+code=cpu" cl0_f">chdoef">false); 34name="L205"> 205 261static int watchdog_nmi_enable(unsigned int 205 257{ 268 struct perf_e">hrtimer *perf_ef">false); 2_timer_fn(struct perf_event *duration = hrtimer, 270 271 t alf">_y setk tommeref">wa?ackup detector */ 2 if (&>222 event-> = 205 293 goto = false); 194 121 waackup detector */ 2 if (duration = 205 287 goto = false); 138 169 perf_">d&>event->< clhwf="+code=wd_hw_attr" clhwf="+chef_ef">false); 269 perf_ref">event->get_sam">duration = 241 151 /*wa function icker functions */ 278 duration = te_pan> te_pan> (perf_f">hrtimer, hrtimer, hrtimer, < class="smes befo_ touboverflow_callback" class="smes befo_ touboveass=f">hrtimer, 194 121 orment"futne b 2 if (&>222 (hrtimer) 287 unsigned long ref="+code=cpu" cl0_f">chdo">duration = ( 228 280static void ( 290 ent"ard0snt"diffee wi 8">ard0sker functions */ 281 if (ref="+code=cpu" cl0_f">chdof">hrtimer) 282 87 unsigned long panic(waaalltou CPUs, fuman wicomconsumeint e hw-PMU dog_if=.\n on cpu %s return; 293 goto = false); 294 } 205 296 wapnree task ief=esf">ormagaillker functions */ 2 if (&>2 if ((ref="+code=cpu" cl0_f">chdoff">hrtimer) 248 duration = ( 289 296 nolker functions */ 2 if ((hrtimer) 292 panic(hrtimer, 2 2 if ((hrtimer) panic(/*wa tion ic oteref">wa\n on cpu %dL215"> 215 282 87 ">hrtimer, 2 else 287 unsigned long ref="+code=cpu"g "f">at_getf">panic(te function :"LOCKUla\n on cpu %dL215"> 215 218 hrtimer, ( 252 return ( 270 271 = 278 hrtimer, duration = = 275 (hrtimer, = 252 0 return; } 289 217static void watchdog_nmmi_enable(unsigned int 205 ) { 1_timer_fn(struct perf_event *duration = hrtimer, 273 1 if ( 282 75 ( 286 struct hrtimer, duration = 157 238 1 219 struct ( 294 } 252 return; } 261static int watchdog_nmi_enable(unsigned int 262static void watchdog_nmmi_enable(unsigned int CONFIG_HARDLrdlockups once */ 157 298watch/>watchdtrouCL27icker functions */ 299/* Calicker functions */ 206#ifdef 157 217static void watch_all"t himer_interrupts"a href="">watch_all"t hiog_nm217<"L205"> 205 ) { 2mi_enable(unsigned int 194 2 if (watchddmer_interrupts"a href="">watchdds="sfg_enabled) { 286 struct watchddmer_interrupts"a href="">watchdds="s">d0 return; 287 struct ( 205 288 87 struct (hrtimer, 2} return; 170} 241 262static void watchd_all"t hiog_nm217<"L205"> 205 266{ (unsigned int 205 2 !a>static void watchdds="sfg_enabled) { 287 struct watchddmer_interrupts"a href="">watchdds="s">d1 return; 238 ( 205 219 87 struct (hrtimer, 294 } 170} 202 283 294 7l/5/watchdog5c#L304" id="L304" class=5line"50ame="L284"> 284 294 , class="srh_sofasrtimer_forwa294"> 294 7l/5/5atchdog5c#L295" id="L295" class=53"> 35fame="L285"> 285 276 (unsigned int

(struct perf_event 2t href="+code=cpu"ps="t haw_get_cpu_vrwa29e5" id="L295" class=3line"39ame="L275"> struct } 2809"+code=get_sample_period"ss="sref">get_ clhwf="+code=wd_hw_attr" clhwf="+che"9p4c#L293" id="L293" c__uspu" class=f">hrti__uspu struct hrtibuif (94 } } dura5ion = watchdds="s">5 5 t alf">5y set5atchdog3c#L211 href="kerne4/watchresix resL296" class=4line"49ame="L286"> 2 5n> even5-5gt; 238 _callback"aet_sasspare_this_cp5erne4/wat5hdog4c#L214" id="L214" c5ass=45ine"41ame="L194"> 194 2 5nsrtimer_f waackup detector */ <5 href="ke5ne4/watchdog4c#L216" id=5L216"5class=4line"41a="kerne4/watchresix resL2968 2t href="+code hrvec_minmaxn; } hrtibuif (94 } } 2 5ned int 205 2 5nple_perio d&>event->< clhwf=5+code=wd_5w_attr" clhwf="+chef_ef"5false51f="kerne4/watchdo" id="L282" class=4line"48ame= get_sam5>dura5u" ree wime="ewatchdog(); ) { 2 5 5 watchdog_nmi5ns */ 266{ 2 5) { 2 5/ 5a href="kerne4/watchdog45#L22652"L277" class=3line"37ame="L217"> 265ERLOCKUP_5ETECTOR"IS_ERLaw_get_cpu5var 157 (unsig_ERLOCKUP5DETECTOR"PTR_ERLaw_get_c5u_var52L27icker functions */ stat5c void watchdds="s">5e=event" 5lass="ssfg_enabled) 5 269 atchdog5c#L300" id="L300" class=5line"50ame="L240"> 294watchdds="s">5 function5 */ hwatchdog3c#L278" id="L278sref">hwatchdog3cog4c#L231"=3line"34ame="L289"> 289watchdds="s">5 /a> 5+code=cpu" cl0_f">chdof"5hrtim5r) hhref="kerne3/watcsref">hhr_hw_attr" clhwf="+=3line"34ame="L289"> 289<#L278" id="L278" class=="_earupts"a href="">watchdds="s">5 s */waaalltou CPUs, 5fuman5wicomconsumein"kerne3/watchdogsref">h7h7hrtimer, (); ) { watchdds="s">5 /a> 5/a> (); watchdds="s">5 5 wapnr5e tas53536" id="L236"kerne3/watchdogdog" class="sre+code=>hrtimer(); ) { watchdds="s">5 RLOCKUP_5href="kerne4/watchdog4c#5237" 5d="L239" class=3line"33ame="L169"> 165PTR_ERLOC5UP_DETECTOR"PTR_ERLaw_ge5_cpu_53L27icker functions */ _, < c> } loame="er_resta_ C_M" id="L294" class=4line"49ame="Ls9f="ke5 5 watchdds="s">5nctions *5 " clid="2" ikerne3/watchdog3cump ois_a> " clid="2" ike" id=e="L290"> 269 */ { ard0ske5de=event"5class="sshdog-t_cpu_var<5a>((); 215 8">ard5 2 5ass=">wat5hdd ( clLOCKUi):e otesup5orWat5(nolLAPIC?)\n on cpu %d">hrtimer<"L238"> 238 2 5="+code=e5ent" class="sshdog-t_cpu5var 125pan class5">watchdd ( clLOCKUi):e"5/*wa 5 tion " class=3line"3name="L155"> 125p 5/watchdog4c#L246" id="L256" cl5ss=4li

e="Loriginal LXR298 b7"> 2 class=3lihttp://sourceforge.net> 28jress/lxr">LXR27 8"ex claa> lxr@#L2ux.no94 lxr.#L2ux.no kindly hosass=b7"class=3lihttp://www.redpill-#L2 28.no">Redpill LL2 28 AS94 28vider of LL2ux/a> glfee w+codo cla_SYMBOserviceBOsi"L291995.