Re: dealing with excessive includes

[Al Viro <viro@ftp.linux.org.uk>]

On Wed, Oct 18, 2006 at 08:04:24AM -0700, Linus Torvalds wrote:
> 
> 
> On Wed, 18 Oct 2006, Al Viro wrote:
> >
> > +#define lock_super(x) do {		\
> > +	struct super_block *sb = x;	\
> > +	get_fs_excl();			\
> > +	mutex_lock(&sb->s_lock);	\
> > +} while(0)
> 
> Don't do this. The "x" passed in may be "sb", and then you end up with 
> bogus code.

*duh*
 
> I think the solution to these kinds of things is either
>  - just bite the bullet, and make it out-of-line. A function call isn't 
>    that expensive, and is sometimes actually cheaper due to I$ issues.
>  - have a separate trivial header file, and only include it for people who 
>    actually need these things (very few files, actually - it's usually 
>    just one file per filesystem)
> 
> In this case, since it's _so_ simple, and since it's _so_ specialized, I 
> think #2 is the right one. Normally, uninlining would be.

Actually, after reading that code I suspect that get_fs_excl() in there
is the wrong thing to do.  Why?  Because the logics is all wrong.

Look what we do under lock_super().  There are two things: ->remount_fs()
and ->write_super().  Plus whatever low-level filesystems are using
lock_super() for.

I would argue that we want to move get_fs_excl() down to the places in
->write_super() that actually want to do something deserving it.  And
to be honest, I'm not at all sure that lock_super() should survive
at upper layers, but that's a longer story...