On Tue, Jan 18, 2005 at 05:54:21PM +0000, Christian Smith wrote:
[...]
> pmap layer on SPARC. Pmap is what implements virtual to physical address
> mapping, and is the location of the bug.

Thanks so much for your post.  This is another of those classics which
should be archived and indexed.  Really improves understanding for us users.

>                |    Process    |
>                |---------------|
>                | Address Space |
> CPU Indepenent | - - (UVM) - - | Virtual Addresses
>                | Memory Objects|
>       ---------|---------------|
>  CPU dependent |     PMAP      |-------------------
>           -----|---------------|
>       Hardware | Physical Mem. | Physical Addresses
>                +---------------+
> 
> As you can see, UVM is pretty entrenched in the workings of the kernel:)

That's for sure!

So, taken together, I should understand this to mean that the NetBSD
developers know where the bug is (in pmap), but don't know what it is?
Or is it that the real bug is in the sun4c hardware itself and no one
has figured out a suitable workaround?

Wasn't I reading somewhere that Sun had released the source to Solaris,
i.e., it's now open-source?  If that's the case, maybe Sun's implementation
of pmap would offer clues as to how to fix the bug.

> I've had some IPX crashes (also sun4c) under heavy load with NetBSD
> 2.0_BETA, but could not pin the problem down to hardware or software. I
> may try reproducing the problem in the future, but for the time being, the
> IPX is retired.

Well, as much as I hate to, it may be wise for someone like me to retire
those IPCs.  As a last resort, I may install Solaris2.6 (at least I think
that version still supported sun4c) to verify that the machines themselves
are okay hardware-wise.

Thanks ALL for your education, suggestions and advice.

-- 
henry nelson
 | day job: | http://yuba.kcn.ne.jp/biorec/nehan/henken.html