Message-ID: <87k0ze7wna.fsf@osv.gnss.ru>
Date:   Wed, 08 Jul 2020 20:18:49 +0300
From:   Sergey Organov <sorganov@...il.com>
To:     Richard Cochran <richardcochran@...il.com>
Cc:     Vladimir Oltean <olteanv@...il.com>, netdev@...r.kernel.org,
        linux-kernel@...r.kernel.org, Fugang Duan <fugang.duan@....com>,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>
Subject: Re: [PATCH  3/5] net: fec: initialize clock with 0 rather than
 current kernel time

Richard Cochran <richardcochran@...il.com> writes:

> On Wed, Jul 08, 2020 at 03:37:29PM +0300, Sergey Organov wrote:
>> Richard Cochran <richardcochran@...il.com> writes:
>> 
>> > On Mon, Jul 06, 2020 at 06:27:21PM +0300, Vladimir Oltean wrote:
>> >> There's no correct answer, I'm afraid. Whatever the default value of the
>> >> clock may be, it's bound to be confusing for some reason, _if_ the
>> >> reason why you're investigating it in the first place is a driver bug.
>> >> Also, I don't really see how your change to use Jan 1st 1970 makes it
>> >> any less confusing.
>> >
>> > +1
>> >
>> > For a PHC, the user of the clock must check the PTP stack's
>> > synchronization flags via the management interface to know the status
>> > of the time signal.
>> 
>> Actually, as I just realized, the right solution for my original problem
>> would rather be adding PTP clock ID that time stamped Ethernet packet to
>> the Ethernet hardware time stamp (see my previous reply as well).
>
> I think you misunderstood my point.  I wasn't commenting on the
> "stacked" MAC/PHY/DSA time stamp issue in the kernel.

I think I did. It's rather that initialization value of PHP clock has
consequences in MAC/PHY/DSA, and there is no way to check any flags
/there/. And it's exactly the place where I needed the info, see
background explanation below.

I understand what your point is, but I try to explain that it's
irrelevant for my particular use-case.

>
> I am talking about the question of whether to initialize the PHC time
> to zero (decades off from TAI) or ktime (likely about 37 seconds off
> from TAI).  It does not really matter, because the user must not guess
> whether the time is valid based on the value.  Instead, the user
> should query the relevant PTP data sets in a "live" online manner.

I'm implementing PTP master clock, and there are no PTP data sets (yet)
in my case, as it's me who will eventually create them.

>
> For example, to tell whether a PHC is synchronized to anything at all,
> you need to check PORT_DATA_SET.portState and probably also
> CURRENT_DATA_SET.offsetFromMaster depending on your needs.

These are fields of PTP stack software that is not running in my case.

>
> In addition, if you care about global time, you need to check:
>
> TIME_PROPERTIES_DATA_SET
>   currentUtcOffset
>   currentUtcOffsetValid
>   ptpTimescale
>   timeTraceable
>   frequencyTraceable

I'm going to become PTP master clock, I already have very precise
estimations of PTP time, I know UTC offset, all this independently from
any PTP stack. Moreover, I've already synchronized PHY clock to this
time scale with +-5ns accuracy, and then I suddenly get somewhat wrong
hardware time stamps for Ethernet packets that I send/receive. You see?

Setting initial value of PTP clock to 0 would allow me to figure what
happens (time stamp coming from /different/ PTP clock) half a day
earlier. Not a big deal, I agree, yet I wanted to help a little bit
somebody who might happen to run into a similar trouble in the future.

Thanks,
-- Sergey

Powered by blists - more mailing lists