Race condition in mb_free_ext()?

[John Baldwin]

On Tuesday 01 March 2005 06:49 pm, Bosko Milekic wrote:
> You know, the more and more we run into these problems specific to how
> reference counting is performed, I keep wondering whether some cleverly
> defined macros for dealing with common reference counting operations
> would be worthwhile.  I'm not saying "introduce a refcount_t type and a
> full-blown abstraction," but some template-like stuff might be useful.
> It took a while just to get the refcount decrement on free path free of
> races on SMP, and that's only in the mbuf code.

Yeah, I have those simple refcount_foo() macros that operate on ints that 
would work for here and things like ucreds.  Being macros, each arch can 
override if they have a better native instruction (xadd on x86, fetchadd on 
ia64, etc.)

> -Bosko
>
> On Tue, Mar 01, 2005 at 03:14:38PM -0800, Kris Kennaway wrote:
> > On Tue, Mar 01, 2005 at 06:04:27PM -0500, Bosko Milekic wrote:
> > > This does not appear to explain the livelock.
> >
> > alc and dwhite tracked it down to a missing volatile causing gcc to
> > mis-optimize the loop:
> >
> > -               cnt = *(m->m_ext.ref_cnt);
> > +               cnt = *(volatile u_int *)(m->m_ext.ref_cnt);
> >
> > I'm currently testing that.
> >
> > Kris

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org