[openib-general] Ordering between PCI config space writes and MMIO reads?

[Jeremy Higdon]

On Tue, Oct 24, 2006 at 06:27:55PM -0500, Jack Steiner wrote:
> On Tue, Oct 24, 2006 at 04:36:32PM -0600, Matthew Wilcox wrote:
> > On Tue, Oct 24, 2006 at 02:51:30PM -0700, Roland Dreier wrote:
> > >  > I think the right way to fix this is to ensure mmio write ordering in
> > >  > the pci_write_config_*() implementations.  Like this.
> > > 
> > > I'm happy to fix this in the PCI core and not force drivers to worry
> > > about this.
> > > 
> > > John, can you confirm that this patch fixes the issue for you?
> > 
> > Hang on.  I wasn't thinking clearly.  mmiowb() only ensures the write
> > has got as far as the shub.  
> 
> I think mmiowb() should work on SN hardware. mmiowb() delays until shub
> reports that all previously issued PIO writes have completed. 
> 
> The processor "mf.a" guarantees "platform acceptance" which on SN means
> that shub has accepted the write - not that it has actually completed (or
> even forwarded anywhere by shub). That makes "mf.a" more-or-less useless
> on SN. However, shub has an additional MMR register (PIO_WRITE_COUNT) that
> counts actual outstanding PIOs.  mmiob() delays until that count goes to
> zero.
> 
> I'll check if there is any additional reordering that can occur AFTER the
> PIO_WRITE_COUNT goes to zero.  If so, it would be at bus level - not in
> shub or routers.

As I understand it, the mmiowb on the shub waits only for the PIO write
to be accepted by the destination node (shub or tio) that the I/O device
is attached to, thus guaranteeing that no reordering will happen within
the NL.

If the PPB can reorder the write, then mmiowb is not sufficient.  You'd
have to do a readback from a chip register (assuming you can trust the
PPB not to reorder reads and writes), or some other work around I haven't
thought of.

jeremy