[Linux-cachefs] [PATCH 2/4] fscache: Fix race in fscache_op_complete() due to split atomic_sub & read

[David Howells]

Andrea Parri <andrea.parri at amarulasolutions.com> wrote:

> > > > > > Fix this by using atomic_sub_return() instead of two calls.
> > > > > 
> > > > > Seems a case for atomic_sub_return_relaxed()... why not?
> > > > 
> > > > Ummm...  In that case, should it be atomic_sub_return_release()?
> > > 
> > > Hard to tell for me: your diff./changelog is all I know about fs-cache
> > > ... (and this suggests -no-, given that atomic_sub() and atomic_read()
> > > provide no ordering...); good question though. ;-)
> > 
> > Yeah, that doesn't mean that it shouldn't be stricter than 'relaxed'.
> > It's kind of like an unlock/release operation, so I think 'release' is
> > probably the minimum requirement.
> 
> Sure.  My point was: those operations are currently not atomic _and_
> they provide no ordering; I think that the above commit message does
> a good work in explaining *why* we need atomicity, but can't say the
> same for the memory-ordering requirement.

Having discussed it with Paul McKenney and thought about it some more, I think
relaxed is probably okay since there isn't a pair of variables that need
ordering.

David