cvs commit: src/sys/sys bus.h kobj.h param.h src/sys/kern subr_bus.c subr_kobj.c

[Bruce Evans]

On Thu, 16 Oct 2003, Doug Rabson wrote:

> On Thu, 2003-10-16 at 12:16, Bruce Evans wrote:
> > On Thu, 16 Oct 2003, Doug Rabson wrote:
> > >     ...
> > >   * Change the kobj method lookup algorithm to one which is SMP-safe. This
> > >     relies only on the constraint that an observer of a sequence of writes
> > >     of pointer-sized values will see exactly one of those values, not a
> > >     mixture of two or more values. This assumption holds for all processors
> > >     which FreeBSD supports.
> >
> > This assumption should be avoided by using atomic_load() (and
> > atomic_store_mumble()).  See a discussion of "atomicity of unlocked
> > reads" last month.  First implement atomic_load().  There is currently
> > only atomic_load_acq_<type>().  "acq" gives acquire semantics which
> > is more than what is needed here and our implementations may do more
> > than what is required anyway for some arches.  "<type>" is part of a
> > bad API.
>
> Without using something like <type>, how can you know how much to read
> (without abusing some kind of GCC language extension)? I don't want to
> use atomic_load_acq_* since I don't care which order values are written
> to the cache.

sizeof(parameter) gives the size (but not the type) if atomic_foo(parameter)
is a macro.  The type isn't really needed in MD macros, especially for
load/store.  The macros for this would be almost as ugly as the ones for
i386 pcpu accesses though.  Using GCC extensions would probably be less ugly.

Bruce