Re: [PATCH 2/2] Introduce __cond_lock_err

[Luc Van Oostenryck]

On Sat, Dec 23, 2017 at 05:06:21AM -0800, Matthew Wilcox wrote:
> On Sat, Dec 23, 2017 at 01:39:11AM -0800, Josh Triplett wrote:
> > +linux-sparse
> 
> Ehh ... we've probably trimmed too much to give linux-sparse a good summary.
> 
> Here're the important lines from my patch:
> 
> +# define __cond_lock_err(x,c)  ((c) ? 1 : ({ __acquire(x); 0; }))
> 
> +       return __cond_lock_err(*ptlp, __follow_pte_pmd(mm, address, start, end,
> +                                                   ptepp, pmdpp, ptlp));
> 
> This is supposed to be "If "c" is an error value, we don't have a lock,
> otherwise we have a lock".  And to translate from linux-speak into
> sparse-speak:
> 
>  # define __acquire(x)  __context__(x,1)
> 
> Josh & Ross pointed out (quite correctly) that code which does something like
> 
> if (foo())
> 	return;
> 
> will work with this, but code that does
> 
> if (foo() < 0)
> 	return;
> 
> will not because we're now returning 1 instead of -ENOMEM (for example).
> 
> So they made the very sensible suggestion that I change the definition
> of __cond_lock to:
> 
> # define __cond_lock_err(x,c)  ((c) ?: ({ __acquire(x); 0; }))
> 
> Unfortunately, when I do that, the context imbalance warning returns.
> As I said below, this is with sparse 0.5.1.

I think this __cond_lock_err() is now OK (but some comment about
how its use is different from __cond_lock() would be welcome).

For the context imbalance, I would really need a concrete example
to be able to help more because it depends heavily on what the
test is and what code is before and after.

If you can point me to a tree, a .config and a specific warning,
I'll be glad to take a look.

-- Luc Van Oostenryck