Date:   Fri, 2 Dec 2022 15:22:11 -0500
From:   Alan Stern <stern@...land.harvard.edu>
To:     Jonas Oberhauser <jonas.oberhauser@...wei.com>
Cc:     Jonas Oberhauser <jonas.oberhauser@...weicloud.com>,
        "paulmck@...nel.org" <paulmck@...nel.org>,
        "parri.andrea@...il.com" <parri.andrea@...il.com>,
        "will@...nel.org" <will@...nel.org>,
        "peterz@...radead.org" <peterz@...radead.org>,
        "boqun.feng@...il.com" <boqun.feng@...il.com>,
        "npiggin@...il.com" <npiggin@...il.com>,
        "dhowells@...hat.com" <dhowells@...hat.com>,
        "j.alglave@....ac.uk" <j.alglave@....ac.uk>,
        "luc.maranget@...ia.fr" <luc.maranget@...ia.fr>,
        "akiyks@...il.com" <akiyks@...il.com>,
        "dlustig@...dia.com" <dlustig@...dia.com>,
        "joel@...lfernandes.org" <joel@...lfernandes.org>,
        "urezki@...il.com" <urezki@...il.com>,
        "quic_neeraju@...cinc.com" <quic_neeraju@...cinc.com>,
        "frederic@...nel.org" <frederic@...nel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] tools: memory-model: Make plain accesses carry
 dependencies

On Fri, Dec 02, 2022 at 05:22:57PM +0000, Jonas Oberhauser wrote:
> > but to me OOTA suggests something more: a value arising as if by 
> > magic rather than as a result of a computation.  In your version of 
> > the litmus test there is WRITE_ONCE(*y, 1), so it's a little 
> > understandable that you could end up with 1 as the final values of x 
> > and y.  But in my version, no values get computed anywhere, so the 
> > final value of x and y might just as easily be 1 or 56789 -- it 
> > literally arises "out of thin air".
> 
> Maybe one can distinguish further between OOTA values (which are 
> arbitrary, not-computed values) and more generally OOTA behaviors.
> 
> How do you feel about examples like the one below:

There's something wrong with this example.

> void *y[2];
> void *x[2] = { (void*)&y[1], (void*)&y[0] };
> 
> P0() {
>     void **t = (void**)(x[0]);

Now t holds a pointer to y[1].

>     *t = (void*)(t-1);

And now y[1] holds a pointer to y[0].

> }
> P1() {
>     void **u = (void**)(x[1]);	

Now u holds a pointer to y[0].

>     *u = (void*)(u+1);

And now y[0] holds a pointer to y[1].

> }
> 
> In this test case the locations x[0] and x[1] exist in the program and 
> are accessed, but their addresses are never (explicitly) taken or 
> stored anywhere.

Although they are dereferened.

> Nevertheless t=&x[1] and u=&x[0] could happen in an appropriately weak 
> memory model (if the data races make you unhappy, you can add relaxed 
> atomic/marked accesses); but not arbitrary values --- if t is not 
> &x[1], it must be &y[1].

I don't see how.  The comments I added above show what values t and u 
must hold, regardless of how the program executes.  The contents of x[] 
never get changed, so there's no question about the values of t and u.

> To me, OOTA suggests simply that the computation can not happen 
> "organically" in a step-by-step way, but can only pop into existence 
> as a whole, "out of thin air" (unless one allows for very aggressive 
> speculation and rollback).

All right, this is more a matter of personal taste and interpretation.  
Is it the computation or the values that pops into existence?  You can 
think of these OOTA computations as arising in a (sort of) ordinary 
step-by-step way, provided you allow loads to read from stores that 
haven't happened yet (a very aggressive form of speculation indeed!).

> In this context I always picture the famous Baron Münchhausen, who 
> pulled himself from mire by his own hair. (Which is an obviously false 
> story because gentlemen at that time were wearing wigs, and a wig 
> could not possibly carry his weight...)

There is a comparable American expression, "pull oneself up by one's 
bootstraps", from which is derived the term "boot" for starting up a 
computer.  :-)

Alan

Powered by blists - more mailing lists