On Thu, Dec 13, 2001 at 04:57:03PM -0500, Jay Estabrook wrote:
> On Thu, Dec 13, 2001 at 10:22:03PM +0100, Kurt Garloff wrote:
> > And how do you handle large requests then? You can't expect to find large
> > blocks of physically contigous memory to do I/O with Linux.
> SCSI drivers typically setup their own S/G list of page-sized or less
> entries; one can (and one used to, before we did the S/G DMA window
> support) simply fill-in the direct-map addresses for each of the SCSI
> S/G entries.
Sure. I happen to maintain trwo SCSI drivers. One of them (tmscsim, driver
for AM53C974 based adapters) even needs an interrupt per page to have the
CPU set up the next segment.
IDE does the same, BTW> The PDR segments also contain a 32bit bus address
and a length field (which often is a page -- if it always was, we would not
have this discussion ... )
> This is fine on machines whose memory can be entirely
> direct-mapped, which is true for MIATA.
> > So disabling S/G makes the performance suck. (For IDE it might still be
> > acceptable. For SCSI, things would already hurt more.)
> No, S/G still takes place, but it depends only on the SCSI HW and not on
> the Alpha S/G DMA window support.
> > My docs say it should be possible to have 6GB. Don't know if that's actually
> > true.
> Yes, theoretically. I think 3GB has been done, in which case one would
> have trouble without S/G DMA window support. Most likely the practical
> limit is 1.5GB.
I think that even up to 4GB should not be a problem, as that's the size of
32bits and the size of the monster window, no?
-- Kurt Garloff <firstname.lastname@example.org> [Eindhoven, NL] Physics: Plasma simulations <K.Garloff@Phys.TUE.NL> [TU Eindhoven, NL] Linux: SCSI, Security <email@example.com> [SuSE Nuernberg, DE] (See mail header or public key servers for PGP2 and GPG public keys.)
This archive was generated by hypermail 2b29 : Sat Dec 15 2001 - 21:00:33 EST