Re: [PATCH 1/2] printk: Store all three timestamps

[Petr Mladek]

On Thu 2020-09-24 10:07:57, Mark Salyzyn wrote:
> I would hope for monotonic_raw, boottime and realtime as being the most
> useful for most situations.
> 
> [TL;DR]
> 
> Currently kernel logs actually uses monotonic_raw (no timing clock
> correction), not monotonic (timing correction).
> 
> Whereas boot (timing clock adjusted, still monotonic) and realtime (timing
> clock _and_ time adjusted, non monotonic), when we try to correlate to user
> space is workable, but we will have troubles correlating monotonic (w.r.t.
> monotonic_raw) clocks if used in user space.
> 
> In Android we have the option of monotonic and realtime only right now for
> the user space logger (which integrates logd, klogd and auditd, the later
> two come from the kernel). I have bugs open to consider boottime, but it is
> blocked on boottime availability from kernel space logging (this change). I
> have another bug to consider switching the logger to monotonic_raw instead
> of monotonic, to make it correlate better with existing kernel logs. But
> alas a lot of resistance for phones switching to monotonic(_raw), the only
> devices that chose monotonic(_raw) is everything else (google glass,
> watches, ...). As such, phones, and the associated developers, will
> continue to want realtime correlated in the kernel logs (yes, this change
> too).
> 
> realtime sucks for the first 10 seconds on Android, since phones generally
> do not get their time correction until then from network resources, and
> many of their rtc clocks are not adjustable, they store a correction factor
> that does not get picked up from user space until userdata is mounted
> (about 20 seconds in). But only kernel developers care about this first
> part of boot, everything after that (and associated correlated kernel
> interactions) are for user space folks.

Thanks a lot for this detailed feedback.

Just to be sure that I understand it correctly. You suggest to store
three timestamps: local_clock(), boot and real clock.

It makes sense to me. I just wonder if there might be any use case
when the adjusted mono clock is needed or preferred over local_clock().

Best Regards,
Petr