[gentoo-dev] New 2.4.18_pre3 (2.4.18_pre3+mjc+XFS) kernel and Athlon bug workaround

[Daniel Robbins]

Hi All,

If you have an Athlon system, be sure this email in its entirety.  A
potential CPU bug fix workaround is covered at the bottom.

I've spent the last few days working on a new kernel for Gentoo Linux. 
This kernel is *excellent* (it's basically the -mjc kernel plus XFS) and
contains tons of security-related enhancements, including:

XFS                                            (from 16 Jan 01 CVS)
Preemptible Kernel Patch (2.4.18-pre3)         (Robert M. Love)
Waitqueue Patch                                (William Lee Irwin III)
Parallel Scanning Patch                        (William Lee Irwin III)
Reverse Mapping Patch #11b                     (Rik van Riel)
IEEE-1394 fixes                                (Andrew Morton)
Mini Low Latency Patch                         (Andrew Morton)
Loopback Deadlock Fixes                        (Andrew Morton)
Read Latency Fixes                             (Andrew Morton)
Truncate Garbage Fixes                         (Andrew Morton)
kGDB                                           (kgdb team)
Zerocopy Pipes                                 (Manfred Spraul)
Pollselect Fixes                               (Manfred Spraul)
Jiffies wraparound fix                         (Tim Schmielau)
RivaFB Software blanking                       (?)
O(1) Scheduler                                 (Ingo Molnar)
-O(1) 'sleep history' comment readability      (me)
TUX linux kernel webserver                     (Ingo Molnar et al)

You can read more about the filesystem read latency fixes at:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0201.0/1654.html

The low-down is that it improves disk IO performance up to 30x (yes,
that's *30 times* in certain situations (generally combined read/write
activity).  And that's just one patch.

This kernel is on Portage now at sys-kernel/linux-sources-2.4.18_pre3.
If you see a _pre4, don't use it.  It will work but it missing nearly
all of these performance enhancing patches, so _pre3 is best.

When testing this kernel, I experienced a problem with my
hardware-accelerated NVIDIA drivers where my system would lock up
whenever I started a *multithreaded* OpenGL application.  I was able to
work around this problem by typing 'export __GL_SINGLE_THREADED="yes"'
before starting the program from the shell.  However, I of course wanted
to be able to use NVIDIA's OpenGL implementation in all its
multithreaded glory.  Fortunately, I was able to reach Terrence Ripperda
from the NVIDIA Linux driver team.  He didn't have a solution but did
mention that there quite a few Athlon CPUs have a particular bug where
the Athlon will corrupt memory if it is using 4Mb page tables with AGP. 
He said that you could tell the kernel to *not* use 4Mb page tables by 
passing the 'mem=nopentium' option to the kernel at boot-time, ie via
GRUB.  I tried this and my problems completely disappeared.  I
immediately contacted some kernel developers, and now Alan Cox and
Terrence are going to try to track down this particular bug and add
support for autodetecting it to the Linux kernel.

Microsoft seems to have discovered this bug, too.  See:
http://support.microsoft.com/default.aspx?scid=kb;en-us;Q270715
Their fix involves changing the default page table size (preventing a
big page table) by tweaking the registry.

So, if you are having stability problems (lock-ups, etc) on an Athlon
system, particularly with 2.4.18_pre3 which seems to make this bug more
prominent, try using the 'mem=nopentium' option which tells the kernel
to effectively work around this CPU bug.  My system is now rock solid
and thanks to the new kernel is also running incredibly fast.

Best Regards,

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