Re: [PATCH] Define wc_wmb, a write barrier for PCI write combining

[Jeremy Higdon <jeremy@sgi.com>]

On Tue, Feb 28, 2006 at 06:20:32PM +0100, Jes Sorensen wrote:
> Roland Dreier wrote:
> >    Jes> Not quite correct as far as I understand it. mmiowb() is
> >    Jes> supposed to guarantee that writes to MMIO space have
> >    Jes> completed before continuing.  That of course covers the
> >    Jes> multi-CPU case, but it should also cover the write-combining
> >    Jes> case.
> >
> >I don't believe this is correct.  mmiowb() does not guarantee that
> >writes have completed -- they may still be pending in a buffer in a
> >bridge somewhere.  The _only_ effect of mmiowb() is to make sure that
> >writes which have been ordered between CPUs using some other mechanism
> >(i.e. a lock) are properly ordered by the rest of the system.  This
> >only has an effect systems like very large ia64 systems, where (as I
> >understand it), writes can pass each other on the way to the PCI bus.
> >In fact, mmiowb() is a NOP on essentially every architecture.
> 
> Hmmmm
> 
> That could be, seems like Jesse agrees that it could all be in the
> pipeline somewhere. Considering Jesse was responsible for mmiowb() I'll
> take his word for it ;-)
> 
> In any case, I'd strongly recommend that any new barrier version is
> clearly documented. The jungle is very dense already ;(


I wonder if wc_wmb() is the best name.

mmiowb expands to memory-mapped I/O write barrier which more or less
describes what it does, whereas wc_wmb expands (I'm guessing) to
write-combine write memory barrier.  But it's for mmio writes.

Also, the wmb() does not actually "guarantee that PCI writes have been
flushed to the bus", at least on IA64.  Even for memory transactions,
it only guarantees ordering on IA64, much like mmiowb does for mmio
transactions.

jeremy