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

From: "Paul E. McKenney" <paulmck@linux.ibm.com>
To: Alan Stern <stern@rowland.harvard.edu>
Cc: Dmitry Vyukov <dvyukov@google.com>,
	LKMM Maintainers -- Akira Yokosawa <akiyks@gmail.com>,
	Andrea Parri <andrea.parri@amarulasolutions.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>,
	Peter Zijlstra <peterz@infradead.org>,
	Will Deacon <will.deacon@arm.com>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: Plain accesses and data races in the Linux Kernel Memory Model
Date: Tue, 15 Jan 2019 07:23:40 -0800	[thread overview]
Message-ID: <20190115152340.GX1215@linux.ibm.com> (raw)
In-Reply-To: <Pine.LNX.4.44L0.1901151002040.1408-100000@iolanthe.rowland.org>

On Tue, Jan 15, 2019 at 10:03:26AM -0500, Alan Stern wrote:
> On Tue, 15 Jan 2019, Dmitry Vyukov wrote:
> 
> > On Tue, Jan 15, 2019 at 12:54 AM Paul E. McKenney <paulmck@linux.ibm.com> wrote:
> > >
> > > On Mon, Jan 14, 2019 at 02:41:49PM -0500, Alan Stern wrote:
> > > > The patch below is my first attempt at adapting the Linux Kernel
> > > > Memory Model to handle plain accesses (i.e., those which aren't
> > > > specially marked as READ_ONCE, WRITE_ONCE, acquire, release,
> > > > read-modify-write, or lock accesses).  This work is based on an
> > > > initial proposal created by Andrea Parri back in December 2017,
> > > > although it has grown a lot since then.
> > >
> > > Hello, Alan,
> > >
> > > Good stuff!!!
> > >
> > > I tried applying this in order to test it against the various litmus
> > > tests, but no joy.  Could you please tell me what commit is this patch
> > > based on?
> > >
> > >                                                         Thanx, Paul
> > >
> > > > The adaptation involves two main aspects: recognizing the ordering
> > > > induced by plain accesses and detecting data races.  They are handled
> > > > separately.  In fact, the code for figuring out the ordering assumes
> > > > there are no data races (the idea being that if a data race is
> > > > present then pretty much anything could happen, so there's no point
> > > > worrying about it -- obviously this will have to be changed if we want
> > > > to cover seqlocks).
> > 
> > Hi Alan,
> > 
> > Is there a mailing list dedicated to this effort? Private messages
> > tend to lost over time, no archive, not possible to send a link or
> > show full history to anybody, etc.
> 
> No specific mailing list.  We've been relying on LKML.
> 
> > Re seqlocks, strictly saying defining races for seqlocks is not
> > necessary. Seqlocks can be expressed without races in C by using
> > relaxed atomic loads within the read critical section. We may consider
> > this option as well.
> 
> That seems like a reasonable approach.

What Alan said!  ;-)

							Thanx, Paul