Re: [kbuild] [djwong-xfs:djwong-wtf 349/351] fs/xfs/xfs_bmap_util.c:1372 xfs_map_free_extent() warn: missing error code 'error'

[Darrick J. Wong]

On Tue, Mar 22, 2022 at 08:47:26AM +0300, Dan Carpenter wrote:
> On Mon, Mar 21, 2022 at 02:59:08PM -0700, Darrick J. Wong wrote:
> > On Mon, Mar 21, 2022 at 10:33:02AM +0300, Dan Carpenter wrote:
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1365  
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1366  	error = xfs_alloc_find_freesp(tp, pag, cursor, end_agbno, &len);
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1367  	if (error)
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1368  		goto out_cancel;
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1369  
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1370  	/* Bail out if the cursor is beyond what we asked for. */
> > > b82670045aab66 Darrick J. Wong 2022-01-06  1371  	if (*cursor >= end_agbno)
> > > b82670045aab66 Darrick J. Wong 2022-01-06 @1372  		goto out_cancel;
> > > 
> > > This looks like it should have an error = -EINVAL;
> > 
> > Nope.  If xfs_alloc_find_freesp moves @cursor goes beyond end_agbno, we
> > want to exit early so that the xfs_map_free_extent caller will return to
> > userspace.
> > 
> > --D
> 
> I'm generally pretty happy with this static checker rule.  Returning
> success on a failure path almost always results if something bad like a
> NULL deref or a use after free.  But false positives are a real risk
> because it's tempting to add an error code to this and introduce a bug.
> 
> Smatch will not print the warning if error is set within 4 lines of the
> goto.
> 	error = 0;
> 	if (*cursor >= end_agbno)
> 		goto out_cancel;

The trouble is, if I do that:

	error = xfs_alloc_find_freesp(...);
	if (error)
		goto out_cancel;

	error = 0;
	if (*cursor >= end_agbno)
		goto out_cancel;

then I'll get patch reviewers and/or tools complaining about setting
error to zero unnecessarily.  Either way we end up with a lot of code
golf for something the compiler will probably remove for us.

Question: Can sparse detect that the if() test involves a comparison
between a non-pointer function argument and a dereferenced pointer
argument?  Would that be sufficient to detect functions that advance a
cursor passed in by the caller and return early when the cursor moves
outside of a range also specified by the caller?

--D

> Another option is that people have started adding comments to these
> blocks in response to the checker warning.
> 
> Or if you had a different idea about how to silence the checker warning
> I can also probably implement that.
> 
> regards,
> dan carpenter
> 
>