[openib-general] 2.6.18-rc5-mm1: drivers/infiniband/hw/amso1100/c2.c compile error

[Roland Dreier]

    Roland> Yes, I agree that's a good plan, especially the
    Roland> documentation part.  However I would argue that what's in
    Roland> drivers/infiniband/hw/mthca/mthca_doorbell.h is
    Roland> legitimate: the driver uses __raw_writeq() when it exists
    Roland> and uses two __raw_writel()s properly serialized with a
    Roland> device-specific lock to get exactly the atomicity it needs
    Roland> on 32-bit archs.

    Bryan> On the off chance that you might be arguing that
    Bryan> mthca_write64 could be a candidate drop-in for writeq on
    Bryan> 32-bit arches:

No, quite the opposite.  I'm arguing that the wrappers in mthca do
legitimately belong in a device driver, since they encapsulate
device-specific knowledge about what serialization suffices when an
atomic __raw_writeq() is not available.

    Bryan> That approach might work on mthca hardware, but it's not
    Bryan> safe in general.  The ipath driver requires a proper
    Bryan> writeq(), for example, because the hardware will quite
    Bryan> legitimately treat 32-bit writes to some registers as
    Bryan> separate accesses, and screw things up royally.

Yes, that's an unfortunate feature of the ipath hardware that
apparently makes it impossible to drive on a generic 32-bit architecture.

So perhaps writeq()/__raw_writeq() need to be defined to generate a
single bus cycle to the extent that makes sense.  Which would mean
that it's not possible to implement on all architectures.

 - R.