Re: arm64: allmodconfig: BUG: KCSAN: data-race in p9_client_cb / p9_client_rpc

From: Marco Elver
Date: Mon Dec 05 2022 - 02:01:17 EST


On Sun, 4 Dec 2022 at 00:08, Dominique Martinet <asmadeus@xxxxxxxxxxxxx> wrote:
>
> Marco Elver wrote on Sat, Dec 03, 2022 at 05:46:46PM +0100:
> > > But I can't really find a problem with what KCSAN complains about --
> > > we are indeed accessing status from two threads without any locks.
> > > Instead of a lock, we're using a barrier so that:
> > > - recv thread/cb: writes to req stuff || write to req status
> > > - p9_client_rpc: reads req status || reads other fields from req
> > >
> > > Which has been working well enough (at least, without the barrier things
> > > blow up quite fast).
> > >
> > > So can I'll just consider this a false positive, but if someone knows
> > > how much one can read into this that'd be appreciated.
> >
> > The barriers only ensure ordering, but not atomicity of the accesses
> > themselves (for one, the compiler is well in its right to transform
> > plain accesses in ways that the concurrent algorithm wasn't designed
> > for). In this case it looks like it's just missing
> > READ_ONCE()/WRITE_ONCE().
>
> Aha! Thanks for this!
>
> I've always believed plain int types accesses are always atomic and the
> only thing to watch for would be compilers reordering instrucions, which
> would be ensured by the barrier in this case, but I guess there are some
> architectures or places where this isn't true?
>
>
> I'm a bit confused though, I can only see five places where wait_event*
> functions use READ_ONCE and I believe they more or less all would
> require such a marker -- I guess non-equality checks might be safe
> (waiting for a value to change from a known value) but if non-atomic
> updates are on the table equality and comparisons checks all would need
> to be decorated with READ_ONCE; afaiu, unlike usespace loops with
> pthread_cond_wait there is nothing protecting the condition itself.
>
> Should I just update the wrapped condition, as below?
>
> - err = wait_event_killable(req->wq, req->status >= REQ_STATUS_RCVD);
> + err = wait_event_killable(req->wq,
> + READ_ONCE(req->status) >= REQ_STATUS_RCVD);

Yes, this looks good!

> The writes all are straightforward, there's all the error paths to
> convert to WRITE_ONCE too but that's not difficult (leaving only the
> init without such a marker); I'll send a patch when you've confirmed the
> read looks good.
> (the other reads are a bit less obvious as some are protected by a lock
> in trans_fd, which should cover all cases of possible concurrent updates
> there as far as I can see, but this mixed model is definitely hard to
> reason with... Well, that's how it was written and I won't ever have time
> to rewrite any of this. Enough ranting.)

If the lock-protected accesses indeed are non-racy, they should be
left unmarked. If some assumption here turns out to be wrong, KCSAN
would (hopefully) tell us one way or another.

Thanks!

-- Marco