Re: [PATCH 1/2] cleanup: Add conditional guard support

[Linus Torvalds]

On Thu, 2 Nov 2023 at 23:30, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Nov 02, 2023 at 03:40:11PM +0100, Oleg Nesterov wrote:
> >
> > To me
> >
> >       guard(rcu);
> >       guard(spinlock, &lock);
> >
> > looks better than
> >
> >       guard(rcu)();
> >       // doesn't match scoped_guard(spinlock, &lock)
> >       guard(spinlock)(&lock);
> >
> > And this will make guard() consistent with scoped_guard().
[...]
> That said; if we were to do this, then something like:
>
> #define __cond_guard(_name, _inst, _fail, args...) \
>         CLASS(_name, _inst)(args); \
>         if (!__guard_ptr(_name)(&_inst)) _fail
>
> #define cond_guard(_name, _fail, args...) \
>         __cond_guard(_name, __UNIQUE_ID(guard), _fail, args)
>
>   cond_guard(spinlock_try, return -EBUSY, &my_lock);
>
> Becomes possible.
>
> Linus, do you like that enough to suffer a flag day patch as proposed by
> Oleg?

I don't find myself caring too much whether we have that "double
grouping" of the guard type-vs-arguments or the "(type, arg...)"
syntax.

I honestly think that "guard(spinlock)(&lock)" makes it more visually
obvious that the first argument is the "type of guard", while
"guard(spinlock, &lock)" makes it look like the two arguments are
somehow at the same level, which they most definitely aren't.

But I also can't find it in myself to care too much about something
that is so purely syntactic, and that I suspect should be abstracted
away anyway to just become "guard_spinlock(&lock)" with a trivial
helper macro.

                Linus