PCI caching question related to mvme5500

Peter Dufault dufault at hda.com
Tue Nov 1 17:14:36 UTC 2005


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1


On Nov 1, 2005, at 11:32 AM, Thomas Doerfler (nt) wrote:

> I have monitored the current discussion. I am not quite sure what you
> actually want to achive with caching the memory area of your PCI card.
>
> Accessing memory over the PCI with single accesses is dead slow, and
> this is most likely what happens, when you copy its content to local
> memory of you MVME5500. Enabling the data cache for that region will
> perform burst accesses, and this will result in a higher troughput,  
> with
> the disadvantage of running into cache coherency problems.

I do a DCBI (Data Cache Block Invalidate) prior to reading the data.   
And yes, I'm reading through the cached region to get the burst access.

>
> Most PCI bridges include some DMA channels. I don't know about the
> MVME5500 PCI bridge, but maybe you can use a built-in DMA channel to
> transfer content efficently from your PCI card to your local memory?

The MVME5500 uses the Marvell GT64260 system controller.  Specs on it  
are only available with an NDA.  I'm reading 16 samples every  
20-25KHz and eventually I'll be reading the data from a custom board  
so this is fine for now.

Maybe Til Straumann has the NDA spec and can package things up for me  
as he did for the VME DMA interface (hint, hint), from recent posts I  
think he has an mvme5500 now.

Peter

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (Darwin)

iD8DBQFDZ6J852YoO31HXPIRAhUOAKCw2uLvRCnotgOKJQo/agTJWYMfXQCgnfSJ
B90jgUu5jozZlrFd9hmNQm8=
=KB7l
-----END PGP SIGNATURE-----



More information about the users mailing list