> >> Almost every report of this sort of thing has come down to faulty boards
> >> or I/O. Almost - thats not every. Can you try and pin down things that make
> >> a difference - eg disabling the PCI optimisation and 4Mb pages. Also if you
> >> have a B stepping 100MHz pentium we need to know that.
>
> Linux should detect and report the B stepping 100MHZ pentium.
> Do it in bright blinking letters followed by a 1 second pause.
As this seems to be easy, it should be done. If there's something
about AMD CPUs it should be done, too.
>
> scream() anyone?
What about writing the message with the faulty command (rep movsb?)
20 times in the video memory and then doing a "clean" panic?
>
> > While hardware problems are not specifically the fault of the linux
> > kernel, people that run other OSes on the same hardware without problems
> > will continue to lay blame on Linux as the culprit. Since Linux's
> > acceptance has as much to do with perception as technical merit, these
> > perceptions have risk. [Not to mention that some pride is at stake, who
> > wants to be outdone by the engineers at Microsoft, even if only in an
> > area as mundane as robustness?]
> >
> > It's great that Linux can and does excel at extracting lots of
> > performance out of our "lowly" Intel boxes, but it seems that this has
> > some price to it. Perhaps, a concerted effort to figure out how flakey
> > hardware misbehaves, and some technique for working around these issues
> > should be considered, if such work is not already underway.
>
> Config option: ASSUME_QUALITY_HARDWARE
>
> The Makefile links hardware.h to quality_hw.h or broken_hw.h.
> These files have #defines for a variety of hardware options.
> When you run 'make config', it only asks for one or the other.
> This gives the newbie an easy way to create a reliable kernel,
> and puts all the options in one file for experts who want to
> customize for a particular machine.
>
> I'm afraid the default should be reliable operation, using a link
> to broken_hw.h. There are too many pieces of junk out there.
> Maybe even assume i386 if broken_hw.h is selected.
386 can be broken, too (was it multiply or divide?)
Ulrich
------------
Ulrich Windl Klinikum der Universitaet Regensburg
Rechenzentrum DV-med Franz-Josef-Strauss-Allee 11
Tel: +49 941 944-5879 D-93053 Regensburg
FAX: +49 941 944-5882
Just imagine my mail address were <Ulrich.Windl@rz.uni.r.de>...