Re: [PATCH] bug: Use xchg() to update WARN_ON_ONCE() static variable

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Oct 15, 2013 at 03:58:06PM -0400, Steven Rostedt wrote:
> The WARN_ON_ONCE() code is to trigger a waring only once when some
> condition happens. But due to the way it is written it is racy.
> 
> 	if (unlikely(condition)) {
> 		if (WARN(!__warned))
> 			__warned = true;
> 	}
> 
> The problem is that multiple CPUs could hit the same warning and
> produce multiple output dumps of the same warning, or an interrupt could
> happen and hit the same warning and do the warning in the middle of a
> previous one, especially since the WARN() does a dump of the current
> stack.
> 
> Even more of a problem, a recent WARN_ON_ONCE() that was in the page
> fault handler triggered and the stack dump of the WARN() caused the
> same WARN_ON_ONCE() get hit again. Since the __warned = true is not
> updated until after the WARN() is completed, each WARN() triggered
> another page fault causing the stack to be filled and crashed the box.
> 
> The point of WARN_ON() is to warn the user and not to crash the box.
> 
> The easy fix is to update the __warned variable with a xchg(). This way
> only one WARN_ON_ONCE() will actually happen, and prevents any issues
> of the WARN() causing the same WARN() to be hit and crash the system.

How about just updating __warned without a cmpxchg. It's not that critical
if the update is not seen immediately to other CPUs. OTOH it's critical
that's it is visible immediately to the current CPU

I mean some warrning can be hard to reproduce and happen to some users
while staying for several kernel releases. If it's repetitive, the xchg
might impact the performance.

I may be overly paranoid, but I think barrier() (so that at least
we don't recurse locally) alone would be better.