Re: Plain accesses and data races in the Linux Kernel Memory Model

From: Alan Stern <stern@rowland.harvard.edu>
To: Andrea Parri <andrea.parri@amarulasolutions.com>
Cc: LKMM Maintainers -- Akira Yokosawa <akiyks@gmail.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	Daniel Lustig <dlustig@nvidia.com>,
	David Howells <dhowells@redhat.com>,
	Jade Alglave <j.alglave@ucl.ac.uk>,
	Luc Maranget <luc.maranget@inria.fr>,
	Nicholas Piggin <npiggin@gmail.com>,
	"Paul E. McKenney" <paulmck@linux.ibm.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Will Deacon <will.deacon@arm.com>,
	Dmitry Vyukov <dvyukov@google.com>,
	<linux-kernel@vger.kernel.org>
Subject: Re: Plain accesses and data races in the Linux Kernel Memory Model
Date: Tue, 22 Jan 2019 11:19:13 -0500 (EST)	[thread overview]
Message-ID: <Pine.LNX.4.44L0.1901221114010.1504-100000@iolanthe.rowland.org> (raw)
In-Reply-To: <20190122154745.GA13659@andrea>

On Tue, 22 Jan 2019, Andrea Parri wrote:

> > @@ -131,7 +159,7 @@ let rec rcu-fence = rcu-gp | srcu-gp |
> >  	(rcu-fence ; rcu-link ; rcu-fence)
> >  
> >  (* rb orders instructions just as pb does *)
> > -let rb = prop ; po ; rcu-fence ; po? ; hb* ; pb*
> > +let rb = prop ; po ; rcu-fence ; po? ; hb* ; pb* ; [marked]
> 
> Testing has revealed some subtle semantics changes for some RCU tests
> _without_ unmarked memory accesses; an example is reported at the end
> of this email.  I suspect that the improvements you mentioned in this
> thread can restore the original semantics but I'm reporting this here
> for further reference.
> 
> With the above definition of 'rb', we're losing links which originate
> or target RCU fences, so that this definition is in fact a relaxation
> w.r.t. the current semantics (even when limiting to marked accesses).
> The test below, for example, is currently forbidden by the LKMM, but
> it becomes allowed with this patch.
> 
> FWIW, I checked that including the RCU fences in 'marked' can restore
> the original semantics of these tests; I'm still not sure whether this
> change can make sense though....
> 
> Thoughts?

Ah, a very good discovery.  I think changing marked to ~plain in a few
places would be a better solution.  Or maybe allowing plain accesses in
those places will also be okay -- it's hard to judge at this point.

> Oh, one last (and unrelated) nit before I forget: IIUC, we used to
> upper-case set names, so I'd also suggest s/marked/Marked, s/plain/Plain
> and similarly for the other sets to be introduced.

Okay, I'll follow that convention.

Alan