On Sat, Dec 15, 2007 at 12:16:59PM +0800, Herbert Xu wrote:
> On Fri, Dec 14, 2007 at 12:02:46PM -0600, Matt Mackall wrote:
> >
> > I added CONFIG_BUG, and I think the current behavior is correct. As
> > you've noticed, we have to evaluate condition, it may have
> > side-effects. And if code does:
> > 
> > 	/* this indicates a driver bug, report and fail gracefully */
> > 	if (WARN_ON(val == NULL))
> > 		return -EFAULT;
> 
> That's exactly the sort of use I had in mind :) I'm actually the
> one who added the ability to use WARN_ON inside an if clause.
> 
> Just as the case of a BUG_ON, a WARN_ON should never occur in
> practice, unless there is a bug which the code is not aware of.

Agreed.

> As such we want it to go away completely if CONFIG_BUG is off.

No. The code as written above should reduce to:

	if (val == NULL)
		return -EFAULT;

If I hadn't wanted to return -EFAULT in this case, I would have just written:

	WARN_ON(val == NULL);

I don't want code that was running safely (ie returning -EFAULT) to
start crashing the system just because I've, say, disabled printk.
That's creating an obnoxious heisenbug.

> > ..we surely want it to continue returning -EFAULT, regardless of
> > whether we log it, no? What use case did you have in mind?
> 
> If you're using it for a scenario which is known to actually
> occur, then some other mechanism should be chosen in place of
> WARN_ON.

Then I kindly submit that you should instead withdraw the code that
allows you to use WARN_ON in a condition in the first place.

Note that Dave Jones is currently poking at making WARN_ON
out-of-line, so you're liable to collide with him.

-- 
Mathematics is the supreme nostalgia of our time.