R1000 status

R1000 status

[Giuseppe Sacco]

Hi all,
reading many web pages I think to understand that R10000 is supported on
IP22 systems and is unsupported on IP32 systems because of its cache.

The page http://www.linux-mips.org/wiki/Silicon_Graphics display a
different result: R10000 on SIGG O2 is supported.

Is there any news or a list of known problems?

Thanks,
Giuseppe

Re: R1000 status

[peter fuerst]

Hello,

yes, the GCC-patches should do the trick. And without a significant
performance decrease, as i was told in the latest success-report (see
http://sgi.sedlacek.biz/).

kind regards


On Sat, 19 May 2007, Giuseppe Sacco wrote:

> Date: Sat, 19 May 2007 08:51:36 +0200 (CEST)
> From: Giuseppe Sacco <giuseppe@eppesuigoccas.homedns.org>
> To: linux-mips@linux-mips.org
> Subject: R1000 status
>
> Hi all,
> reading many web pages I think to understand that R10000 is supported on
> IP22 systems and is unsupported on IP32 systems because of its cache.
>
> The page http://www.linux-mips.org/wiki/Silicon_Graphics display a
> different result: R10000 on SIGG O2 is supported.
>
> Is there any news or a list of known problems?
>
> Thanks,
> Giuseppe
>
>
>
>

Re: R1000 status

[Kumba]

Giuseppe Sacco wrote:
> Hi all,
> reading many web pages I think to understand that R10000 is supported on
> IP22 systems and is unsupported on IP32 systems because of its cache.
> 
> The page http://www.linux-mips.org/wiki/Silicon_Graphics display a
> different result: R10000 on SIGG O2 is supported.
> 
> Is there any news or a list of known problems?

R10000 on IP22 actually refers to the IP28 system, Indigo2 R10000 Impact.  And 
it does work if you use the latest patches, but YMMV.

R10000 on IP32 is a different story.  I used to maintain a patch that was 
essentially the IP28 patch minus IP28-specific bits to the kernel, and built 
with an R10K-equipped gcc toolchain, produced a kernel that would run for a 
decent amount of time, however, using the scsi disk or the network module could 
lock the machine up.  You also got a gazillion CRIME CPU errors on the main console.

I'll have to rebuild the patch again and see how it handles nowadays, but I 
believe that, regardless, enhancements need to be made to aic7xxx to protect 
against the non-cache cohenecy of the machine, and something equivalent to the 
meth module (network) as well.  That, or someone implements the "Juice" 
capability of the system.


--Kumba

-- 
Gentoo/MIPS Team Lead

"Such is oft the course of deeds that move the wheels of the world: small hands 
do them because they must, while the eyes of the great are elsewhere."  --Elrond

Re: R1000 status

[Kumba]

peter fuerst wrote:
> 
> Hello,
> 
> yes, the GCC-patches should do the trick. And without a significant
> performance decrease, as i was told in the latest success-report (see
> http://sgi.sedlacek.biz/).
> 
> kind regards

Interesting, I'm curious to see this guy's patches to see what he does different 
for the IP32 R10K stuff.  Whether it's the same thing I did (mod IP28 patches 
and build w/ gcc cache barriers on load and store ops), or if he went an extra 
step and added code to protect the scsi and networking drivers.


--Kumba

-- 
Gentoo/MIPS Team Lead

"Such is oft the course of deeds that move the wheels of the world: small hands 
do them because they must, while the eyes of the great are elsewhere."  --Elrond