Re: [PATCH 0/2 v6] printk, Add monotonic and real printk timestamps

[Thomas Gleixner]

On Thu, 10 Mar 2016, Petr Mladek wrote:
> On Tue 2016-03-08 06:03:24, Prarit Bhargava wrote:
> > 
> > 
> > On 03/08/2016 02:59 AM, Thomas Gleixner wrote:
> > > On Tue, 23 Feb 2016, Prarit Bhargava wrote:
> > > 
> > >> This patchset adds monotonic and real printk timestamps.  The first patch
> > >> changes CONFIG_PRINT_TIME from a bool to an int to allow for the additional
> > >> timestamps that are added in patch 2.
> > >>
> > >> Changes from v6: Petr Mladek pointed out that the current patch
> > >> fails to indicate to userspace programs which timestamp is being used.
> > > 
> > > How is that solved?
> > 
> > Hi Thomas,
> > 
> > Userspace programs can now look at /sys/modules/printk/parameters/time which
> > will contain [0-3] for the timestamp clock.
> 
> But it includes only the current setting that is valid only for
> messages printed with this setting. The ring buffer might include
> different messages produced with different setting.
> 
> I suggest to look how dmesg handles the time stamp. I wonder how it
> converts the relative time into an absolute one. I wonder if you
> could convert all timestamps to the relative format, so that you
> do not need to change all userspace tools at all.

  ctime = system_boot_time + dmesg_stamp

The reason why you see wrong values here is that dmesg_stamp is not based on
clock monotonic. So if printk uses clock monotonic, dmesg does not have to be
changed.

Though you have to know, that the above is a linearized history because
system_boot_time is 

  clock_monotonic_now() + boot_offset

So you can't observe stuff like leap seconds, settimeofday or other things
which result in a non linear behaviour of clock realtime.

Thanks,

	tglx