Message-Id: <20181214002043.GP4170@linux.ibm.com>
Date:   Thu, 13 Dec 2018 16:20:43 -0800
From:   "Paul E. McKenney" <paulmck@...ux.ibm.com>
To:     Alan Stern <stern@...land.harvard.edu>
Cc:     David Goldblatt <davidtgoldblatt@...il.com>,
        mathieu.desnoyers@...icios.com,
        Florian Weimer <fweimer@...hat.com>, triegel@...hat.com,
        libc-alpha@...rceware.org, andrea.parri@...rulasolutions.com,
        will.deacon@....com, peterz@...radead.org, boqun.feng@...il.com,
        npiggin@...il.com, dhowells@...hat.com, j.alglave@....ac.uk,
        luc.maranget@...ia.fr, akiyks@...il.com, dlustig@...dia.com,
        linux-arch@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] Linux: Implement membarrier function

On Thu, Dec 13, 2018 at 10:49:49AM -0500, Alan Stern wrote:
> On Wed, 12 Dec 2018, Paul E. McKenney wrote:
> 
> > > Well, what are you trying to accomplish?  Do you want to find an 
> > > argument similar to the one I posted for the 6-CPU test to show that 
> > > this test should be forbidden?
> > 
> > I am trying to check odd corner cases.  Your sys_membarrier() model
> > is quite nice and certainly fits nicely with the rest of the model,
> > but where I come from, that is actually reason for suspicion.  ;-)
> > 
> > All kidding aside, your argument for the 6-CPU test was extremely
> > valuable, as it showed me a way to think of that test from an
> > implementation viewpoint.  Then the question is whether or not that
> > viewpoint actually matches the model, which seems to be the case thus far.
> 
> It should, since I formulated the reasoning behind that viewpoint 
> directly from the model.  The basic idea is this:
> 
> 	By induction, show that whenever we have A ->rcu-fence B then
> 	anything po-before A executes before anything po-after B, and
> 	furthermore, any write which propagates to A's CPU before A
> 	executes will propagate to every CPU before B finishes (i.e.,
> 	before anything po-after B executes).
> 
> 	Using this, show that whenever X ->rb Y holds then X must
> 	execute before Y.
> 
> That's what the 6-CPU argument did.  In that litmus test we have
> mb2 ->rcu-fence mb23, Rc ->rb Re, mb1 ->rcu-fence mb14, Rb ->rb Rf,
> mb0 ->rcu-fence mb05, and lastly Ra ->rb Ra.  The last one is what 
> shows that the test is forbidden.

I really am not trying to be difficult.  Well, no more difficult than
I normally am, anyway.  Which admittedly isn't saying much.  ;-)

> > A good next step would be to automatically generate random tests along
> > with an automatically generated prediction, like I did for RCU a few
> > years back.  I should be able to generalize my time-based cheat for RCU to
> > also cover SRCU, though sys_membarrier() will require a bit more thought.
> > (The time-based cheat was to have fixed duration RCU grace periods and
> > RCU read-side critical sections, with the grace period duration being
> > slightly longer than that of the critical sections.  The number of
> > processes is of course limited by the chosen durations, but that limit
> > can easily be made insanely large.)
> 
> Imagine that each sys_membarrier call takes a fixed duration and each 
> other instruction takes slightly less (the idea being that each 
> instruction is a critical section).  Instructions can be reordered 
> (although not across a sys_membarrier call), but no matter how the 
> reordering is done, the result is disallowed.

It gets a bit trickier with interleavings of different combinations
of RCU, SRCU, and sys_membarrier().  Yes, your cat code very elegantly
sorts this out, but my goal is to be able to explain a given example
to someone.

> > I guess that I still haven't gotten over being a bit surprised that the
> > RCU counting rule also applies to sys_membarrier().  ;-)
> 
> Why not?  They are both synchronization mechanisms with heavy-weight
> write sides and light-weight read sides, and most importantly, they
> provide the same Guarantee.

True, but I do feel the need to poke at it.

The zero-size sys_membarrier() read-side critical sections do make
things act a bit differently, for example, interchanging the accesses
in an RCU read-side critical section has no effect, while doing so in
a sys_membarrier() reader can cause the result to be allowed.  One key
point is that everything before the end of a read-side critical section
of any type is ordered before any later grace period of that same type,
and vice versa.

This is why reordering accesses matters for sys_membarrier() readers but
not for RCU and SRCU readers -- in the case of RCU and SRCU readers,
the accesses are inside the read-side critical section, while for
sys_membarrier() readers, the read-side critical sections don't have
an inside.  So yes, ordering also matters in the case of SRCU and
RCU readers for accesses outside of the read-side critical sections.
The reason sys_membarrier() seems surprising to me isn't because it is
any different in theoretical structure, but rather because the practice
is to put RCU and SRCU read-side accesses inside a read-side critical
sections, which is impossible for sys_membarrier().

The other thing that took some time to get used to is the possibility
of long delays during sys_membarrier() execution, allowing significant
execution and reordering between different CPUs' IPIs.  This was key
to my understanding of the six-process example, and probably needs to
be clearly called out, including in an example or two.

The interleaving restrictions are straightforward for me, but the
fixed-time approach does have some interesting cross-talk potential
between sys_membarrier() and RCU read-side critical sections whose
accesses have been reversed.  I don't believe that it is possible to
leverage this "order the other guy's read-side critical sections" effect
in the general case, but I could be missing something.

If you are claiming that I am worrying unnecessarily, you are probably
right.  But if I didn't worry unnecessarily, RCU wouldn't work at all!  ;-)

							Thanx, Paul

Powered by blists - more mailing lists