Re: [PATCH] x86: Remove readq()/writeq() on 32-bit

[Roland Dreier <rdreier@cisco.com>]

 > This removal patch is completely pointless, because it moves us
 > backwards to the point where we had a bunch of drivers defining it.

No, the current kernel still requires drivers to define it anyway,
because there are tons of 32-bit architectures that are not x86.

And more than that, centralizing the definition makes the API much more
dangerous for driver authors.

 > At least the networking drivers I messed with (until 11/2008) were
 > always fine with a non-atomic readq.

The commit to niu I keep citing (e23a59e1, "niu: Fix readq
implementation when architecture does not provide one.") shows that
drivers need to take care.  Now, the x86 implementation would happen to
work for that hardware, but eg drivers/infiniband/hw/amso1100 defines
readq with the opposite order -- whether that's required or just an
arbitrary choice, I don't know.  And drivers/infiniband/hw/mthca has
some uses of __raw_writeq() that only work if no other CPU accesses to
the same page can happen between the two halves, so it adds a per-page
spinlock for 32-bit architectures.  etc.

 - R.