dmesg: adjust timestamps according to suspended time
Subsystems (2)
PATCH 1/1 dmesg: adjust timestamps according to suspended time
Timestamps in kernel log comes from monotonic clocksource which does not tick when system suspended. Suspended time easily sums into hours and days rendering human readable timestamps in dmesg useless.
Adjusting timestamps accouring to current delta between boottime and monotonic clocksources produces accurate timestamps for messages printed since last resume. Which are supposed to be most interesting.
--- include/monotonic.h | 2 ++ lib/monotonic.c | 12 ++++++++++++ sys-utils/dmesg.1 | 2 ++ sys-utils/dmesg.c | 22 +++++++++++++++++----- 4 files changed, 33 insertions(+), 5 deletions(-)
include/monotonic.h +2 -0
@@ -15,6 +15,8 @@extern int get_boot_time (struct timeval *boot_time); +extern time_t get_suspended_time (void ); +extern int gettime_monotonic (struct timeval *tv);#endif /* UTIL_LINUX_MONOTONIC_H */
lib/monotonic.c +12 -0
@@ -48,6 +48,18 @@ int get_boot_time(struct timeval *boot_time)#endif } +time_t get_suspended_time (void ) +{ +#if defined(CLOCK_BOOTTIME ) && defined(CLOCK_MONOTONIC ) +struct timespec boot, mono; + +if (clock_gettime (CLOCK_BOOTTIME , &boot) ==0 && +clock_gettime (CLOCK_MONOTONIC , &mono) ==0 ) +return boot.tv_sec - mono.tv_sec; +#endif +return 0 ; +} +int gettime_monotonic (struct timeval *tv) {#ifdef CLOCK_MONOTONIC
sys-utils/dmesg.1 +2 -0
@@ -161,6 +161,8 @@ source used for the logs is .B not updated after system .BR SUSPEND / RESUME . +Timestamps are adjusted according to current delta between boottime and monotonic +clocks, this works only for messages printed after last resume. .IP "\fB\-t\fR, \fB\-\-notime\fR" Do not print kernel's timestamps. .IP "\fB\-\-time\-format\fR \fIformat\fR"
sys-utils/dmesg.c +17 -5
@@ -169,6 +169,7 @@ struct dmesg_control {struct timeval lasttime;/* last printed timestamp */ struct tm lasttm;/* last localtime */ struct timeval boot_time;/* system boot time */ +time_t suspended_time;/* time spent in suspeneded state */ int action;/* SYSLOG_ACTION_* */ int method;/* DMESG_METHOD_* */
@@ -824,7 +825,7 @@ static struct tm *record_localtime(struct dmesg_control *ctl,struct dmesg_record *rec, struct tm *tm) { - time_t t = ctl->boot_time.tv_sec + rec->tv.tv_sec; + time_t t = ctl->boot_time.tv_sec + ctl->suspended_time + rec->tv.tv_sec;return localtime_r (&t, tm); }
@@ -852,7 +853,7 @@ static char *iso_8601_time(struct dmesg_control *ctl, struct dmesg_record *rec, char *buf, size_t bufsz) {struct timeval tv = { - .tv_sec = ctl->boot_time.tv_sec + rec->tv.tv_sec, + .tv_sec = ctl->boot_time.tv_sec + ctl->suspended_time + rec->tv.tv_sec, .tv_usec = rec->tv.tv_usec };
@@ -1301,8 +1302,16 @@ static inline int dmesg_get_boot_time(struct timeval *tv) return get_boot_time(tv); } + +static inline time_t dmesg_get_suspended_time (void ) +{ +if (getenv ("DMESG_TEST_BOOTIME" )) +return 0 ; +return get_suspended_time (); +}#else # define dmesg_get_boot_time get_boot_time +# define dmesg_get_suspended_time get_suspended_time#endif int main (int argc,char *argv[])
@@ -1499,9 +1508,12 @@ int main(int argc, char *argv[])if ((is_timefmt (&ctl,RELTIME ) ||is_timefmt (&ctl,CTIME ) || -is_timefmt (&ctl,ISO8601 )) - &&dmesg_get_boot_time (&ctl.boot_time) !=0 ) - ctl.time_fmt =DMESG_TIMEFTM_NONE ; +is_timefmt (&ctl,ISO8601 ))) { +if (dmesg_get_boot_time (&ctl.boot_time) !=0 ) + ctl.time_fmt =DMESG_TIMEFTM_NONE ; +else + ctl.suspended_time =dmesg_get_suspended_time (); + }if (delta)switch (ctl.time_fmt) {
Karel Zak
maintainerQuoted 7 lines · Konstantin Khlebnikov
Timestamps in kernel log comes from monotonic clocksource which does not tick when system suspended. Suspended time easily sums into hours and days rendering human readable timestamps in dmesg useless.
Adjusting timestamps accouring to current delta between boottime and monotonic clocksources produces accurate timestamps for messages printed since last resume. Which are supposed to be most interesting.
It's definitely better than the current broken timestamps, but the real and final solution is to have exact information about system suspends.
It would be enough to maintain in kernel memory a simple log with
<bootime> <monotonic> <state_change>
and export this info by /proc/suspendlog, after that we can all re-count /dev/kmsg timestamps to something useful.
Karel
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com
Quoted 9 lines · Karel Zak
Quoted 7 lines · Konstantin Khlebnikov
Timestamps in kernel log comes from monotonic clocksource which does not tick when system suspended. Suspended time easily sums into hours and days rendering human readable timestamps in dmesg useless.
Adjusting timestamps accouring to current delta between boottime and monotonic clocksources produces accurate timestamps for messages printed since last resume. Which are supposed to be most interesting.
Quoted 7 lines
It's definitely better than the current broken timestamps, but the real and final solution is to have exact information about system suspends.
It would be enough to maintain in kernel memory a simple log with
<bootime> <monotonic> <state_change>
and export this info by/proc/suspendlog, after that we can all re-count/dev/kmsgtimestamps to something useful.
Boottime or real time could be simply printed into kernel log at suspend and resume. So demsg could detect current offset while reading.
But for older kernel dmesg still needs guessing like this.
Quoted 4 lines
Karel
Karel Zak
maintainerQuoted 13 lines · Konstantin Khlebnikov
Quoted 12 lines · Karel Zak
Quoted 7 lines · Konstantin Khlebnikov
Timestamps in kernel log comes from monotonic clocksource which does not tick when system suspended. Suspended time easily sums into hours and days rendering human readable timestamps in dmesg useless.
Adjusting timestamps accouring to current delta between boottime and monotonic clocksources produces accurate timestamps for messages printed since last resume. Which are supposed to be most interesting.
It's definitely better than the current broken timestamps, but the real and final solution is to have exact information about system suspends.
Quoted 7 lines
Quoted 4 lines
It would be enough to maintain in kernel memory a simple log with
<bootime> <monotonic> <state_change>
and export this info by/proc/suspendlog, after that we can all re-count/dev/kmsgtimestamps to something useful.Boottime or real time could be simply printed into kernel log at suspend and resume. So demsg could detect current offset while reading.
Yes, but not sure if this is the most robust way (dmesg --clear will remove this info) and I guess the suspendlog can be useful independently on kmsg.
Karel
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com
Quoted 5 lines · Konstantin Khlebnikov
include/monotonic.h | 2 ++ lib/monotonic.c | 12 ++++++++++++ sys-utils/dmesg.1 | 2 ++ sys-utils/dmesg.c | 22 +++++++++++++++++----- 4 files changed, 33 insertions(+), 5 deletions(-)
Applied, thanks.
Signature · 2 lines
Karel Zak <[email protected]> http://karelzak.blogspot.com