Re: 2.6.18-rc5-mm1: drivers/infiniband/hw/amso1100/c2.c compile error

[Roland Dreier <rdreier@cisco.com>]

    Roland> My understanding is that __raw_writeq() is like writeq()
    Roland> except not strongly ordered and without the byte-swap on
    Roland> big-endian architectures.  The __raw_writeX() variants are
    Roland> convenient to avoid having to write inefficient code like
    Roland> writel(swab32(foo), ...) when talking to a PCI device that
    Roland> wants big-endian data.  Without the raw variant, you end
    Roland> up with a double swap on big-endian architectures.

Oh, I left one other thing out: writeq() and __raw_writeq() shold be
atomic in the sense that no other transactions should be able to get
onto the IO bus in the middle -- so implementing writeq() as two
writel()s in a row is not allowed

    Andrew> OK.  Can we please stop hacking around this in drivers and

    Andrew> a) work out what it's supposed to do

    Andrew> b) document that (Documentation/DocBook/deviceiobook.tmpl
    Andrew> or code comment or whatever)

    Andrew> c) tell arch maintainers?

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

It's an open question what drivers that don't actually need atomicity
but just want a convenient way to write 64 bits at time should do.

 - R.