Re: TG3 data corruption (TSO ?)

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> > semantics. At least what is implemented currently on PowerPC is the
> > __raw_* versions which not only have no barriers at all (they don't even
> > order between MMIOs, for example, readl might cross writel), and do no
> > endian swap. Quite a mess of semantics if you ask me... Then there has
> 
> __writel/__readl seems more in keeping for just not locking.

Not locking... you mean not ordering I suppose. Ok, so the question is
no ordering at all (that is even between MMIO read/writes, thus a
__readl can cross a __writel), or just no ordering between MMIO and
cacheable storage ?

It's an important difference and both have their use. For example, on
PowerPC, if I completely remove barriers, I get the first semantic and I
get the ability to write combine on non-cacheable storage as a benefit
(provided we add an ioremap_wc or such, as the Guarded bit we set on
normal non-cacheable space does also prevent write combining on most
implementations). However, if I keep at least ordering between MMIOs,
then I leave an eieio in there, which is not nearly as expensive than a
full sync but will not order cacheable cs. non-cacheable. However, it
will also prevent write combine as far as I remember.

Ben.