Re: PCI-Arbiter Extension

[Tom Keaveny <tak@core.rose.hp.com>]

Hmmm...

Would it be possible to run all of the non-bridge REQ#/GNT# pairs to
a standalone arbiter.  This arbiter would then communicate to the
bridge's embedded arbiter via one of the original REG#/GNT# pairs, on
behalf of the other agents. In this way, the original bridge only 
has one GNT# line of interest.

Dealing with the bridge's desire to access the bus is probably a bit
sticky (effectively an internal grant that can change on the fly),
and would likely vary according to the bridge..

==
tom keaveny
hewlett packard co.


> 
> It is impossible unless the original PCI arbiter understands that a cascaded 
> arbiter is to be used.  Since arbitration is pipelined there are conditions in 
> which two devices can believe that they own the bus.
> 
> If the primary arbiter knows that a second level is supported, then the first 
> arbiter must not deassert one GNT# and assert a different GNT# on the same 
> clock. (WHich the spec allows while the bus is active.)  The first level arbiter
> must ALWAYS cause a dead clock between the deassertion of one GNT# and the 
> assertion of the next GNT#.  If the arbiter never deasserts a GNT# until FRAME# 
> is asserted, this can cause a hand system when a master removes its REQ# because
> it changed it mind.  (THis can occur when the master enable bit is cleared.)
> 
> Many have tried this with pals and such but all have failed because the first 
> level arbiter did not account for a second level arbiter.
> 
> Norm
tlY