On Wed, 21 Dec 2005 17:02:55 +0100
Simon Richter <[email protected]> wrote:
> > the proposal actually had a fully-working patch attached :)
>
> Ah, didn't see that, as I just skimmed over the web archive page you
> linked to, which has no link to the actual patch (or I'm too stupid to
> find it).
right.. the link was to 0/6 of the patchset, which is
actually only the introduction. real patch was in subsequent
messages.
> > In my code, the first rtc that register is bound
> > to /proc/driver/rtc and /dev/rtc (if those interfaces
> > are compiled in, as they are all selectable).
>
> It would be good to have a way to change which clock is the "primary"
> one from userspace later (userspace because this is clearly site policy).
If I'm not wrong, the RTC is usually queried at bootup
and written to on shutdown. If NTP mode is active,
it is also written every 11 minutes.
So my intention was to emulate that interface as a starting
point. Then we can update the userspace utilities (hwclock)
to let the user choose which clock he want to use.
I guess /proc/driver/rtc will be deprecated sooner or
later. The /dev/rtc interface only supports one clock.
It can either be extended to have /dev/rtcX or we
can extend the sysfs one to allow clock updating.
NTP mode could then be adjusted to update one or more
of the rtcs. Maybe each RTC could have an attribute
(let's say /sys/class/rtc/rtcX/ntp) which tells the
kernel whether to update it or not.
This way we will not have a primary clock anymore.
> > You have full control of which functions you will provide
> > to the upper layer. Obivously if you try to set the
> > time on a read-only rtc, you will get an error.
>
> Sure. I was thinking of the question which error that should be.
-EPERM ? -EACCESS? :)
--
Best regards,
Alessandro Zummo,
Tower Technologies - Turin, Italy
http://www.towertech.it
Hello,
Alessandro Zummo schrieb:
>>It would be good to have a way to change which clock is the "primary"
>>one from userspace later (userspace because this is clearly site policy).
> If I'm not wrong, the RTC is usually queried at bootup
> and written to on shutdown. If NTP mode is active,
> it is also written every 11 minutes.
A good ntpd will adjust the speed rather than write to the clock; the
ntpd shipped by most distributions can already handle multiple time sources.
I'm thinking of the case where a computer is not attached to a network
but needs accurate tim; in this case I'd give it a battery powered RTC
and a time signal receiver. As most time signals are low-bandwidth, they
may not carry full time information in each tick so it may take several
minutes to fully synchronize. In this case I'd like to use the battery
backed up clock first and switch later on when synchronized.
> I guess /proc/driver/rtc will be deprecated sooner or
> later. The /dev/rtc interface only supports one clock.
> It can either be extended to have /dev/rtcX or we
> can extend the sysfs one to allow clock updating.
/dev is the way to go IMO. As far as I've understood sysfs, it carries
meta information about devices and drivers only, the actual
communication then happens through device nodes still.
> NTP mode could then be adjusted to update one or more
> of the rtcs. Maybe each RTC could have an attribute
> (let's say /sys/class/rtc/rtcX/ntp) which tells the
> kernel whether to update it or not.
That's entirely a userspace thing. What the userspace needs to know from
the kernel is whether the clock is writable and whether its speed can be
adjusted.
> -EPERM ? -EACCESS? :)
-EIO or -ENOSYS would also be possible options.
Simon
On Thu, 22 Dec 2005 00:18:33 +0100
Simon Richter <[email protected]> wrote:
> A good ntpd will adjust the speed rather than write to the clock; the
> ntpd shipped by most distributions can already handle multiple time sources.
Yes, but there's the kernel who writes to the clock,
for example http://lxr.linux.no/source/arch/arm/kernel/time.c?a=arm#L103 .
> > later. The /dev/rtc interface only supports one clock.
> > It can either be extended to have /dev/rtcX or we
> > can extend the sysfs one to allow clock updating.
>
> /dev is the way to go IMO. As far as I've understood sysfs, it carries
> meta information about devices and drivers only, the actual
> communication then happens through device nodes still.
Ok. We can use dynamic device numbers and go for the /dev
interface.
>
> > NTP mode could then be adjusted to update one or more
> > of the rtcs. Maybe each RTC could have an attribute
> > (let's say /sys/class/rtc/rtcX/ntp) which tells the
> > kernel whether to update it or not.
>
> That's entirely a userspace thing. What the userspace needs to know from
> the kernel is whether the clock is writable and whether its speed can be
> adjusted.
agreed. there are are also some variables of interest in
http://lxr.linux.no/source/include/linux/timex.h?a=arm#L188
some of them may be usefully exported in sysfs.
--
Best regards,
Alessandro Zummo,
Tower Technologies - Turin, Italy
http://www.towertech.it