On Thu, Oct 24, 2019 at 04:17:11PM +0200, Marco Elver wrote:
> On Thu, 24 Oct 2019 at 14:28, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> >
> > On Thu, Oct 17, 2019 at 04:13:01PM +0200, Marco Elver wrote:
> > > Since seqlocks in the Linux kernel do not require the use of marked
> > > atomic accesses in critical sections, we teach KCSAN to assume such
> > > accesses are atomic. KCSAN currently also pretends that writes to
> > > `sequence` are atomic, although currently plain writes are used (their
> > > corresponding reads are READ_ONCE).
> > >
> > > Further, to avoid false positives in the absence of clear ending of a
> > > seqlock reader critical section (only when using the raw interface),
> > > KCSAN assumes a fixed number of accesses after start of a seqlock
> > > critical section are atomic.
> >
> > Do we have many examples where there's not a clear end to a seqlock
> > sequence? Or are there just a handful?
> >
> > If there aren't that many, I wonder if we can make it mandatory to have
> > an explicit end, or to add some helper for those patterns so that we can
> > reliably hook them.
>
> In an ideal world, all usage of seqlocks would be via seqlock_t, which
> follows a somewhat saner usage, where we already do normal begin/end
> markings -- with subtle exception to readers needing to be flat atomic
> regions, e.g. because usage like this:
> - fs/namespace.c:__legitimize_mnt - unbalanced read_seqretry
> - fs/dcache.c:d_walk - unbalanced need_seqretry
>
> But anything directly accessing seqcount_t seems to be unpredictable.
> Filtering for usage of read_seqcount_retry not following 'do { .. }
> while (read_seqcount_retry(..));' (although even the ones in while
> loops aren't necessarily predictable):
>
> $ git grep 'read_seqcount_retry' | grep -Ev 'seqlock.h|Doc|\* ' | grep
> -v 'while ('
> => about 1/3 of the total read_seqcount_retry usage.
>
> Just looking at fs/namei.c, I would conclude that it'd be a pretty
> daunting task to prescribe and migrate to an interface that forces
> clear begin/end.
>
> Which is why I concluded that for now, it is probably better to make
> KCSAN play well with the existing code.
Thanks for the detailed explanation, it's very helpful.
That all sounds reasonable to me -- could you fold some of that into the
commit message?
Thanks,
Mark.
