[CentOS] [OT] Memory Models and Multi/Virtual-Cores -- WAS: 4.0 -> 4.1 update failing

Sat Jun 25 01:35:20 UTC 2005
Peter Arremann <loony at loonybin.org>

On Friday 24 June 2005 21:14, Bryan J. Smith wrote:
> On Fri, 2005-06-24 at 20:16 -0400, Peter Arremann wrote:
> > Ok - so the PAE36 mechanism is the same, the 36bit addresses are the same
> > just the address ranges that are reserved for special use are different?
>
> Ugh, sigh, you're thinking like a programmer.  ;->  You have to separate
> the logical (programmer) from the physical (engineer) concepts.
Sorry I guess. Don't know how that happened - went to a electrical engineering 
school, specialized in micro electronics :-) 

> > That would require an additional level of page translation though or a
> > really substential redesign - so its in my opinion unlikely that they
> > do...
>
> Again, thinking like a programmer.  ;->
*whistles innocently* 

> The Athlon was _already_ designed for 40-bit EV6 _physically_ and
> eventual "Long Mode" programmatically.  The TLB in the Athlon is _not_
> designed for GTL, but EV6.  It only works like a GTL when so
> commanded.  ;->
Ok, now I have to ask even more questions :-) If the physical bus is EV6, the 
internals are EV6 and the programming model doesn't matter as you said before 
- then why would the TLB on the athlon emulate a GTL? 

> > > Intel GTL is only capable of 32-bit.
> > GTL (P5) yes - but that doesn't have PAE36 either... But we're talking
> > PAE here - so GTL+ aera...
>
> I meant GTL and all derivatives ... GTL+, AGTL+, etc...
Then you're clearly wrong calling it a 32bit addressed bus... 

> > The address pins are even labeled up to A35 Download any of the PPro and
> > newer spec sheets (except mobile cpus that don't support PAE). The
> > Pentium II 300Mhz Specs is a good example. Quote from page 75: "The
> > Address signals define a 2^36-byte physical memory address space."
> > document id 24365702.pdf
>
> Of course!  Because if it didn't, it wouldn't be able to page in beyond
> 4GiB.  @-ppp  Yes, but the TLB compatibility and other issues are
> involved, things that GTL and all derivatives _never_ addressed.
>
> In a nutshell, Intel PAE36 processors on even AGTL+ are _incapable_ of
> the combination of both _physical_ and _logical_ addressing to do
> anything but paging above 4GiB.  That's my point.
Define paging then - I can access >4GB on an EMT64 - It just lacks and IO mmu 
to do coordinate dma from a 32bit device to a memory region > 4GB... but if I 
just do a memory access, then there is no issue - no paging nessecary... 

> AMD solved the problem by using EV6, which is _nothing_ like GTL-based
> busses.  It goes far deeper than you realize.
they solved the 32bit IO issue with by morphing their agpgart into a 
minimalistic io mmu ...

Peter.