Re: [PATCH 7/7] kvmtool: 9p: refactor rel_to_abs()

[Will Deacon <will.deacon@arm.com>]

On Fri, Nov 18, 2016 at 04:33:07PM +0100, G. Campana wrote:
> On 11/17/2016 01:20 PM, Will Deacon wrote:
> > On Thu, Nov 10, 2016 at 04:18:54PM +0100, G. Campana wrote:
> >> On 08/11/2016 03:38, Will Deacon wrote:
> >>> On Tue, Oct 18, 2016 at 06:03:05PM +0200, G. Campana wrote:
> >>>> @@ -614,7 +618,6 @@ static void virtio_p9_readdir(struct p9_dev *p9dev,
> >>>>  	struct stat st;
> >>>>  	struct p9_fid *fid;
> >>>>  	struct dirent *dent;
> >>>> -	char full_path[PATH_MAX];
> >>>>  	u64 offset, old_offset;
> >>>>  
> >>>>  	rcount = 0;
> >>>> @@ -645,11 +648,8 @@ static void virtio_p9_readdir(struct p9_dev *p9dev,
> >>>>  			break;
> >>>>  		}
> >>>>  		old_offset = dent->d_off;
> >>>> -		if (rel_to_abs(p9dev, dent->d_name, full_path, sizeof(full_path)) != 0) {
> >>>> -			errno = ENAMETOOLONG;
> >>>> -			goto err_out;
> >>>> -		}
> >>>> -		lstat(full_path, &st);
> >>>> +		if (stat_rel(p9dev, dent->d_name, &st) != 0)
> >>>> +			memset(&st, -1, sizeof(st));
> >>>
> >>> Why the memset, and not goto err_out?
> >>>
> >> Because the user may not be allowed to stat some entries in a directory
> >> and it shouldn't make readdir() fail.
> > 
> > Ok, but is memsetting to -1 really the right thing to do? This gets
> > "converted" into a p9_qid_t, which will then look pretty strange (path
> > and version will be set to 0xff, type will be set to P9_QTDIR).
> > 
> Before this patch, st was either uninitialized or invalid if lstat
> failed, hence the memset call which doesn't break the logic of this
> function. I only tried to fix vulnerabilities in this patch series, and
> I think this issue deserves a separate patch. What do you think?

Well, how about just skipping entries where the stat failed and continuing
around the loop. Does that work?

Will