Re: >Re: [RFC] should VM_BUG_ON(cond) really evaluate cond

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 28 octobre 2011 à 04:37 -0700, Linus Torvalds a écrit :
> On Thu, Oct 27, 2011 at 9:43 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> > The only requirement of atomic_read() is that it must return value
> > before or after an atomic_write(), not a garbled value.
> 
> The problem is that gcc *can* return a garbled value.
> 
> > In fact, if a compiler is stupid enough to issue two reads on following
> > code :
> 
> The compiler really *can* be that "stupid". Except the code tends to
> look like this:
> 
>    int value = atomic_read(&atomic_var);
>    if (value > 10)
>      return;
>    .. do something with value ..
> 
> and gcc may decide - under register pressure, and in the absense of a
> 'volatile' - to read 'value' first once for that "> 10" check, and
> then it drops the registers and instead of saving it on the stack
> frame, it can decide to re-load it from atomic_var.
> 
> IOW, "value" could be two or more different values: one value when
> testing, and *another* value in "do something with value".
> 
> This is why we have "ACCESS_ONCE()".
> 
> Whether atomics guarantee ACCESS_ONCE() semantics or not is not
> entirely clear. But afaik, there is no way to tell gcc "access at
> *most* once, and never ever reload".
> 

What you describe is true for non atomic variables as well, its not part
of the atomic_ops documented semantic.

And we do use ACCESS_ONCE() on the rare cases we need to make sure no
reload is done.

RCU use makes this implied (ACCESS_ONCE() being done in
rcu_dereference()), so we dont have many raw ACCESS_ONCE() in our code.

int value = ACCESS_ONCE(atomic_read(&atomic_var));
if (value > 10)
     return;
.. do something with value ..

One of such rare use is explained in commit f1987257
(tcp: protect sysctl_tcp_cookie_size reads)

Since its a bit ugly, I suggested :

int value = atomic_read_once(&atomic_var);
if (value > 10)
     return;
    .. do something with value ..

I dont know, it seems the right way, but yes it might break things.

We can take the otherway and patch thousand atomic_read() to
atomic_read_stable(), its safer but very boring :)