On Thu, 2003-01-30 at 15:50, Andrea Arcangeli wrote:
> On Thu, Jan 30, 2003 at 03:44:38PM -0800, Andrew Morton wrote:
> > Stephen Hemminger wrote:
> > > 
> > > This is an update to the earlier frlock.
> > > 
> > 
> > Sorry, but I have lost track of what version is what.  Please
> > let me get my current act together and then prepare diffs
> > against (or new versions of) that.
> > 
> > You appear to have not noticed my earlier suggestions wrt
> > coding tweaks and inefficiencies in the new implementation.
> > 
> > - SEQ_INIT and seq_init can go away.
> > 
> > - do seq_write_begin/end need wmb(), or mb()?  Probably, we
> >   should just remove these functions altogether.

Since nothing uses them yet, yes, just more to go wrong.


> > 	+static inline int seq_read_end(const seqcounter_t *s, unsigned iv)
> > 	+{
> > 	+       mb();
> > 	+       return (s->counter != iv) || (iv & 1);
> > 	+}
> > 
> >   So the barriers changed _again_!  Could we please at least
> >   get Richard Henderson and Andrea to agree that this is the
> >   right way to do it?

That was actually a typo. Should be rmb().


> the right way is the one used by x86-64 vgettimeofday and
> i_size_read/write in my tree (and frlock in my tree too for x86
> gettimeofday)
> 
> that is pure rmb() in read_lock and pure wmb() in write_lock
> 
> never mb()
> 
> The only place where mb() could be somehow interesting is the
> write_begin/end but it's mostly a theorical interest, and we both think
> that write_begin/end is pointless, since the lock part is useless for
> them, and in turn write_begin/end aren't that clean anyways.
> 

Rather than splitting it into two pieces, counter and lock+counter;
go back to one structure, ince no place is using just the counter alone.
I will merge all this back tomorrow.
	

-- 
Stephen Hemminger <shemminger@osdl.org>
Open Source Devlopment Lab