Re: [PATCH] Make WARN_ON/WARN_ON_ONCE no-ops when CONFIG_BUG is off

[Matt Mackall <mpm@selenic.com>]

On Sat, Dec 15, 2007 at 02:52:11PM +0800, Herbert Xu wrote:
> On Sat, Dec 15, 2007 at 02:04:49PM +0800, Herbert Xu wrote:
> >
> > I suppose we'll have to either introduce a new primitive or just
> > go back to using BUG_ON.
> 
> Let's do the conservative thing and add a new primitive.
> 
> [PATCH] Added BARF_ON/BARF_ON_ONCE
> 
> The description of CONFIG_BUG clearly states that both BUG and
> WARN_ON may be skipped.  However, our actual implementation still
> checks the condition on WARN_ON if it's used as part of an if
> statement or such.
> 
> I tried to compile it away but Matt Mackall pointed out that
> people may already be using it in ways that were not intended.
> In particular, it may have been used for conditions that are
> thought to be possible.
> 
> So this patch adds a new pair of macros BARF_ON and BARF_ON_ONCE
> that are clearly marked as such to prevent abuse.  The intended
> usage model is identical to WARN_ON except that they should only
> be used for conditions that are thought to be impossible.  In
> other words, only use them to replace BUG_ON's that may be called
> by an IRQ handler or within locks.

Egads, this just makes the whole bad BUG_ON vs WARN_ON situation
worse. This adds yet another non-distinction, because the difference
between BUG_ON and WARN_ON has absolutely nothing to do with the
(perceived or actual) probability of the condition.

We've got plenty of evidence that BUG_ON and WARN_ON conditions -do-
happen. So choosing between BUG_ON and WARN_ON is choosing whether a
box should be forcibly killed or not and nothing more.

You've clearly got a use-case in mind where WARN_ON + !CONFIG_BUG
isn't doing what you want, please share it with us.

-- 
Mathematics is the supreme nostalgia of our time.