Re: [PATCH v5 00/18] Rework READ_ONCE() to improve codegen

From: Will Deacon
Date: Tue May 12 2020 - 17:15:03 EST


On Tue, May 12, 2020 at 09:07:55PM +0200, Peter Zijlstra wrote:
> On Tue, May 12, 2020 at 07:53:00PM +0200, Marco Elver wrote:
> > I just ran a bunch of KCSAN tests. While this series alone would have
> > passed the tests, there appears to be a problem with
> > __READ_ONCE/__WRITE_ONCE. I think they should already be using
> > 'data_race()', as otherwise we will get lots of false positives in
> > future.
> >
> > I noticed this when testing -tip/locking/kcsan, which breaks
> > unfortunately, because I see a bunch of spurious data races with
> > arch_atomic_{read,set} because "locking/atomics: Flip fallbacks and
> > instrumentation" changed them to use __READ_ONCE()/__WRITE_ONCE().
> > From what I see, the intent was to not double-instrument,
> > unfortunately they are still double-instrumented because
> > __READ_ONCE/__WRITE_ONCE doesn't hide the access from KCSAN (nor KASAN
> > actually). I don't think we can use __no_sanitize_or_inline for the
> > arch_ functions, because we really want them to be __always_inline
> > (also to avoid calls to these functions in uaccess regions, which
> > objtool would notice).
> >
> > I think the easiest way to resolve this is to wrap the accesses in
> > __*_ONCE with data_race().
>
> But we can't... because I need arch_atomic_*() and __READ_ONCE() to not
> call out to _ANYTHING_.
>
> Sadly, because the compilers are 'broken' that whole __no_sanitize thing
> didn't work, but I'll be moving a whole bunch of code into .c files with
> all the sanitizers killed dead. And we'll be validating it'll not be
> calling out to anything.

Hmm, I may have just run into this problem too. I'm using clang 11.0.1,
but even if I do something like:

unsigned long __no_sanitize_or_inline foo(unsigned long *p)
{
return READ_ONCE_NOCHECK(*p);
}

then I /still/ get calls to __tcsan_func_{entry,exit} emitted by the
compiler. Marco -- how do you turn this thing off?!

I'm also not particularly fond of treating __{READ,WRITE}ONCE() as "atomic",
since they're allowed to tear and I think callers should probably either be
using data_race() explicitly or disabling instrumentation (assuming that's
possible).

Will