Re: [libc-alpha] Re: PATCH: Fix ll/sc for mips

[Andreas Schwab <schwab at suse dot de>]

Hiroyuki Machida <machida@sm.sony.co.jp> writes:

|> From: Kaz Kylheku <kaz@ashi.footprints.net>
|> Subject: Re: [libc-alpha] Re: PATCH: Fix ll/sc for mips
|> Date: Thu, 31 Jan 2002 20:02:25 -0800 (PST)
|> 
|> > On Fri, 1 Feb 2002, Hiroyuki Machida wrote:
|> > > Please note that "sc" may fail even if nobody write the
|> > > variable. (See P.211 "8.4.2 Load-Linked/Sotre-Conditional" of "See 
|> > > MIPS RUN" for more detail.) 
|> > > So, after your patch applied, compare_and_swap() may fail, even if
|> > > *p is equal to oldval.
|> > 
|> > I can't think of anything that will break because of this, as long
|> > as the compare_and_swap eventually succeeds on some subsequent trial.
|> > If the atomic operation has to abort for some reason other than *p being
|> > unequal to oldval, that should be cool.
|> 
|> I mean that this patch breaks the spec of compare_and_swap().
|> 
|> In most case, this patch may works as Kaz said. If this patch have
|> no side-effect to any application, it's ok to apply the patch. But
|> we can't know how to use compare_and_swap() in all aplications in a
|> whole world. So we have to follow the spec.  

There is no way to find out anything about intermediate values of *p when
compare_and_swap returns zero.  The value of *p can change anytime, even
if it only was different from oldval just at the time compare_and_swap did
the comparison.  So there is zero chance that a spurious failure of
compare_and_swap breaks anything.

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE GmbH, Deutschherrnstr. 15-19, D-90429 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."