uninitialized variables [Was: svn commit: r365445 - head/sys/cam/mmc]

Mark Johnston markj at freebsd.org
Wed Sep 9 13:44:36 UTC 2020


On Wed, Sep 09, 2020 at 08:49:01AM +0300, Andriy Gapon wrote:
> On 08/09/2020 15:48, Mark Johnston wrote:
> > I observed the same thing recently as well: the compiler catches
> > uninitialized variables only in simple cases.  In my case, any uses of
> > goto within the function seemed to silence the warning, even if they
> > appeared after the uninitialized reference.
> 
> I am running a kernel build now with this addition (for clang):
> CWARNEXTRA+=   -Wconditional-uninitialized -Wno-error-conditional-uninitialized
> 
> It produces a ton of warnings.
> Some of them are probably false positives, but some look quite reasonable.

It has a lot of trouble with code patterns of the form:

	for (i = 0; i < 100; i++) {
		val = foo();
	}
	if (val != 0) /* may be uninitialized!!1 */
		bar();

or

	if (foo == bar)
		val = baz();
	<some other stuff>
	if (foo == bar && val == 3)
		<some stuff>

The second example makes some sense to me since it's hard to prove that
foo == bar will not change between the first and second evaluations.

> E.g.:
> sys/cam/cam_periph.c:314:19: warning: variable 'p_drv' may be uninitialized when
> used here [-Wconditional-uninitialized]
>                 TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
> 
> Indeed, there is a conditional 'goto failure' before a first assignment to p_drv
> and the line is after the label.  So, maybe the situation is impossible, but it
> is reasonable to warn about it.
>
> But the number of false positives (and "possible but impossible" situations) is
> too overwhelming.

Yeah.  I looked at maybe 30 warnings (out of hundreds) this morning
and they were all false positives.  KMSAN will provide a new tool for
finding such bugs, but they will only be detected at runtime.


More information about the svn-src-head mailing list