[gentoo-dev] Interview with AMD reg. Large Pages Bug

[gentoo-dev] Interview with AMD reg. Large Pages Bug

[Nero24 (Planet 3DNow!)]

Hello Chris,

regarding the Large Pages Bug with Windows 2000 and Linux there are
different version out, whether the bug is fixed in Palomino or not. So we
decided to stop rumoring and asked AMD public relations manager Jan Gütter.
What he said you can read at Planet 3DNow!
www.planet3dnow.de/cgi-bin/newspub/viewnews.cgi?category=1&id=1011807957

Bye

Nero24

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Nero24 - Newseditor, Hardware-Redakteur & Administrator im Forum

 Email: nero24@planet3dnow.de ; nero24@gmx.li
 ICQ: #74295042
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Planet 3DNow! - Alles über 3DNow!, AMD, Athlon & Co

 URL: http://www.planet3dnow.de
 Chat: IRCNet / #planet3dnow
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Re: [gentoo-dev] Interview with AMD reg. Large Pages Bug

[Daniel Robbins]

On Wed, 2002-01-23 at 11:57, Nero24 (Planet 3DNow!) wrote:
> Hello Chris,
> 
> regarding the Large Pages Bug with Windows 2000 and Linux there are
> different version out, whether the bug is fixed in Palomino or not. So we
> decided to stop rumoring and asked AMD public relations manager Jan Gütter.
> What he said you can read at Planet 3DNow!
> www.planet3dnow.de/cgi-bin/newspub/viewnews.cgi?category=1&id=1011807957

The INVPLG CPU bug is not the particular issue being discussed here
(yes, this has all been very confusing).  The issue we are talking about
is a subtle interaction between speculative writes, extended paging and
the GART (not a CPU bug, this was a misunderstanding).  This quirk
results in corruption of AGP and system memory, and it's really the
responsibility of the Linux kernel and drivers to make sure that this
issue doesn't happen.  See my post at
http://www.geocrawler.com/lists/3/Linux/35/75/7626960/ which contains an
update including an official explantion from AMD of a problem that can
happen and may be happening to Linux people.  Follow the thread and
you'll see that it's a real issue.  It is a very interesting problem.

The Linux kernel developers are currently investigating this speculative
write/GART interaction to see if it could affect Linux users.  So far,
the consensus seems to be that it *does* affect Linux AGP users and
developers are currently trying to find a solution that doesn't not hurt
performance as "mem=nopentium" (turning off extended translation) does.

Best Regards,

-- 
Daniel Robbins                                  <drobbins@gentoo.org>
Chief Architect/President                       http://www.gentoo.org 
Gentoo Technologies, Inc.