Re: Adding plain accesses and detecting data races in the LKMM

From: "Paul E. McKenney" <paulmck@linux.ibm.com>
To: Andrea Parri <andrea.parri@amarulasolutions.com>
Cc: Alan Stern <stern@rowland.harvard.edu>,
	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>,
	Peter Zijlstra <peterz@infradead.org>,
	Will Deacon <will.deacon@arm.com>,
	Daniel Kroening <kroening@cs.ox.ac.uk>,
	Kernel development list <linux-kernel@vger.kernel.org>
Subject: Re: Adding plain accesses and detecting data races in the LKMM
Date: Mon, 15 Apr 2019 06:35:35 -0700	[thread overview]
Message-ID: <20190415133535.GU14111@linux.ibm.com> (raw)
In-Reply-To: <20190413213938.GA4371@andrea>

On Sat, Apr 13, 2019 at 11:39:38PM +0200, Andrea Parri wrote:
> On Tue, Apr 09, 2019 at 08:01:32AM -0700, Paul E. McKenney wrote:
> > On Tue, Apr 09, 2019 at 03:36:18AM +0200, Andrea Parri wrote:
> > > > > The formula was more along the line of "do not assume either of these
> > > > > cases to hold; use barrier() is you need an unconditional barrier..."
> > > > > AFAICT, all current implementations of smp_mb__{before,after}_atomic()
> > > > > provides a compiler barrier with either barrier() or "memory" clobber.
> > > > 
> > > > Well, we have two reasonable choices: Say that 
> > > > smp_mb__{before,after}_atomic will always provide a compiler barrier, 
> > > > or don't say this.  I see no point in saying that the combination of 
> > > > Before-atomic followed by RMW provides a barrier.
> > > 
> > > ;-/ I'm fine with the first choice. I don't see how the second choice
> > > (this proposal/patch) would be consistent with some documentation and
> > > with the current implementations; for example,
> > > 
> > > 1) Documentation/atomic_t.txt says:
> > > 
> > > Thus:
> > > 
> > >   atomic_fetch_add();
> > > 
> > > is equivalent to:
> > > 
> > >   smp_mb__before_atomic();
> > >   atomic_fetch_add_relaxed();
> > >   smp_mb__after_atomic();
> > > 
> > > [...]
> > > 
> > > 2) Some implementations of the _relaxed() variants do not provide any
> > > compiler barrier currently.
> > 
> > But don't all implementations of smp_mb__before_atomic() and
> > smp_mb__after_atomic() currently supply a compiler barrier?
> 
> Yes, AFAICS, all implementations of smp_mb__{before,after}_atomic() currently
> supply a compiler barrier.
> 
> Nevertheless, there's a difference between:  (1) Specify that these barriers
> supply a compiler barrier,  (2) Specify that (certain) combinations of these
> barriers and RMWs supply a compiler barrier, and  (3) This patch...  ;-)
> 
> FWIW, I'm not aware of current/informal documentation following (the arguably
> simpler but slightly stronger) (1).  But again (amending my last remark): (1)
> and (2) both make sense to me.

Another question is "should the kernel permit smp_mb__{before,after}*()
anywhere other than immediately before or after the primitive being
strengthened?"

Thoughts?

							Thanx, Paul