Re: per_cpu fixes

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Thu, 10 Jul 2003 11:15:25 -0700 (PDT), Linus Torvalds <torvalds@osdl.org> said:

  Linus> On Thu, 10 Jul 2003, David Mosberger wrote:
  >> 
  >> You mean there would be three primitives:
  >> 
  >> (1) get value from a per-CPU variable
  >> (2) set value of a per-CPU variable
  >> (3) get the (canonical) address of a per-CPU variable

  Linus> Argh.

  Linus> We'd better have the rule that if there are any virtual
  Linus> caches or other issues, then the "canonical address" had
  Linus> better be the _only_ address (or at least any virtual
  Linus> remapping has to be done in such a way that it never causes
  Linus> aliasing or other performance problems with the canonical
  Linus> address).

  Linus> This is already turning fairly ugly, and I just don't want to
  Linus> see even more ugly rules like "you can't mix direct accesses
  Linus> with pointer accesses"

Yes, Rusty's proposal (as I think I summarized above) would do exactly
that: the only address that you'll ever see for a per-CPU variable
will be the canonical one.  The get/set macros can use an alias on
platforms where this is more efficient, but the alias will never be
visible outside the macros.

	--david