Re: [PATCH v3 46/57] perf: Simplify pmu_dev_alloc()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jun 13, 2023 at 09:50:28AM +0200, Greg KH wrote:

> > DEFINE_FRERE(class_destroy, struct class *, if (!IS_ERR_OR_NULL(_T)) class_destroy(_T))
> 
> Nit, as class_destroy() handles this type of check within it, it can be
> even simpler:
> 	DEFINE_FREE(class_destroy, struct class *, class_destroy(_T));

Note that that means there will be an unconditional call to
class_destroy() in the success path. As long as that is never a hot-path
this should be fine I suppose, but it is something Linus pointed out
earlier.

> or would that be:
> 	DEFINE_CLASS(class_destroy, struct class *, class_destroy(_T));

Has a slightly different syntax per the comment I did do write :-)

DEFINE_CLASS(class, struct class *,
	     if (!IS_ERR_OR_NULL(_T)) class_destroy(_T),
	     class_create(cname), const char *name)

static int __init misc_init(void)
{
	struct proc_dir_entry *ret __free(remove_proc) =
		proc_create_seq("misc", 0, NULL, &misc_seq_ops);

	CLASS(class, c)("misc");
	if (IS_ERR(c))
		return PTR_ERR(c);

	if (register_chrdev(MISC_MAJOR, "misc", &misc_fops))
		return -EIO;

	c->devnode = misc_devnode;

	misc_class = no_free_ptr(c);
	no_free_ptr(ret);

	return 0;
}

The no_free_ptr() should work with CLASS(), but I'm not sure that's
recommended, lots of un-explored terretory here :-)

Similarly I suppose you could do something like:

DEFINE_CLASS(proc_dir, struct proc_dir_entry *,
	     proc_remove(_T), proc_create(pname, mode, parent, proc_ops),
	     const char *pname, umode_t mode, struct proc_dir_entry *parent,
	     const struct proc_ops *proc_ops)

EXTEND_CLASS(proc_dir, _seq, proc_create_seq(pname, mode, parent, ops, state_size, data),
	     const char *pname, umode_t mode, struct proc_dir_entry *parent,
	     const struct seq_operations *ops, unsigned int state_size, void *data)

EXTEND_CLASS(proc_dir, _seq_private, .....)

(urgh, C really needs better forwarding support)

Then you could write it something like:

static int __init misc_init(void)
{
	CLASS(proc_dir_seq, ret)("misc", 0, NULL, &misc_seq_ops);

	CLASS(class, c)("misc");
	if (IS_ERR(c))
		return PTR_ERR(c);

	if (register_chrdev(MISC_MAJOR, "misc", &misc_fops))
		return -EIO;

	c->devnode = misc_devnode;

	misc_class = no_free_ptr(c);
	no_free_ptr(ret);

	return 0;
}

Is what what we want?

(also, perhaps I should prefix the macro arguments with an '_', as is
you can't use 'name' as a constructor argument because the thing would
expand weird)

> I have a ton of future patches coming that does a bunch of
> class_create/destroy changes that would be made a LOT simpler with this
> patchset, and I really don't want to have to hit the same codepaths
> twice if at all possible.
> 
> So what's the odds this can be reasonable enough to get into 6.5-rc1 so
> we can rely on it there?

That's one for Linus I suppose.. the only remaining issue I still have
is the no_free_*() naming. One suggestion that made sense had it called
take_*().

All we really need are the first 4 patches to land; thereafter we can
gradually start converting things.