Re: [RFC] High resolution time support using x86 TSC

[Colin D Bennett]

On Thu, 03 Jul 2008 20:52:53 +0200
Marco Gerards <address@hidden> wrote:

> Hi Colin,
> 
> Colin D Bennett <address@hidden> writes:
> 
> > +   * kern/i386/linuxbios/init.c (grub_get_time_ms):
> > +   Define grub_get_time_ms() to always return 0.  This should
> > be fixed
> > +   when RTC functionality is implemented.
> > +   (grub_time_init): Define this required function as a
> > no-op. Inserted
> > +   a comment to remind us delete this function when proper
> > time support 
> > +   is added to i386-linuxbios.
> > +
> > +   * kern/main.c (grub_main): Call grub_time_init() right
> > after
> > +   grub_machine_init().
> 
> I think this should go into grub_machine_init?  Why didn't you just
> add it there?

Originally, I had grub_machine_init() call grub_tsc_calibrate(), but
these are comments Vesa made to me about that code:

Vesa Jääskeläinen <address@hidden> wrote
on Mon, 16 Jun 2008 20:34:54 +0300:

> Colin D Bennett wrote:
> > This week I implemented high resolution timer support using the x86
> > TSC (see attached patch <grub_tsc_2008-06-10.patch>).  It is
> > calibrated at GRUB startup using the RTC as a reference. The core
> > TSC function is ``grub_get_tsc()`` -- corresponding to
> > ``grub_get_rtc()``, but returning a uint64 value representing the
> > number of CPU cycles elapsed since boot.
> > 
> > In most situations, you will want to use ``grub_get_time_ms()`` to
> > get the system time in milliseconds based on the TSC value.  The
> > calibration function ``grub_tsc_calibrate()``, calculates the
> > proper scale factor and absolute offset so that the millisecond
> > value represents real time.
> > 
> > The ``grub_get_time_ms()`` function is implemented for non-x86
> > platforms to simply call ``grub_get_time_ms_generic()`` (defined in
> > kern/misc.c), which uses the RTC to get the time in milliseconds.  
> 
> We would probably want to leave that generic out from kernel, and let 
> every platform either use this generic code or implement their own 
> mechanism to do the job. Perhaps we should make own folder for
> generic stuff that can be included for arch specific build if there
> is no better replacement.
> 
> How about calling function like grub_time_init() which would then be 
> platform specific? Then platform can implement whatever way to
> calibrate (if needed) as long as it provides this grub_get_time_ms()
> function (also being platform specific). This would make
> initialization non-specific to arch while leaving more room for
> implementation.

Therefore, I thought this was the right way to do it.  Do you want me
to instead call grub_time_init() from grub_machine_init()?

Regards,
Colin

signature.asc
Description: PGP signature