[dm-devel] [PATCH 09/14] libmultipath: check whether mp->features is NULl in, assemble_map

[Benjamin Marzinski]

On Tue, Sep 08, 2020 at 06:35:45PM +0200, Martin Wilck wrote:
> On Tue, 2020-09-08 at 10:45 -0500, Benjamin Marzinski wrote:
> > On Mon, Sep 07, 2020 at 08:21:28PM +0800, lixiaokeng wrote:
> > > > > @@ -86,9 +86,12 @@ assemble_map (struct multipath * mp, char *
> > > > > params, int len)
> > > > >  	    get_linux_version_code() < KERNEL_VERSION(4, 3, 0))
> > > > >  		add_feature(&mp->features, retain_hwhandler);
> > > > > 
> > > > > -	f = STRDUP(mp->features);
> > > > 
> > > > clearly strdup()ing without checking if mp->features NULL is
> > > > incorrect.
> > > > However, I'm not sure that we need to fail if mp->features is
> > > > NULL.
> > > > instead, int the APPEND call, we could use the gcc ternary
> > > > operator
> > > > extension
> > > > 
> > > > (mp->features)?: "0"
> > > > 
> > > > to use "0" if mp->features is NULL.
> > > > 
> > > > Also, have you seen this actually occur?  Or is this just a
> > > > theoretical
> > > > issue that you've found from reading the code.  It seems like
> > > > setup_map() will always call select_features() before calling
> > > > assemble_map(), which should mean that mp->features will always
> > > > be set
> > > > in this case. Perhaps I'm missing something here.
> > > > 
> > > > -Ben
> > > > 
> > > Hi Ben,
> > >   This just a theoretical issue and I did not see it. But it's not
> > > necessary
> > > to call strdup. In your opinion, need multipath be checked?  I will
> > > make new
> > > patch with your suggestion.
> > 
> > Since we don't believe it's possible for mp->features (or mp-
> > >hwhandler)
> > to be set to NULL here, it makes sense to print an error if it is
> > NULL.
> > So, I guess my suggestion would be to print an error message if
> > mp->features or mp->hwhandler are NULL, but to assemble the map
> > anyway,
> > using the default value of "0" if they are NULL. That's how
> > assemble_map() currently handles failures in add_feature().
> > add_feature() will print an error, but assemble_map() will go ahead
> > with
> > assembling the map.
> > 
> > I'm willing to be convinced that there is a better solution, however.
> 
> What about this:
> 
> assemble_map() is only called from setup_map(), which sets mp->features 
> in select_features(). So what we should do is check for NULL after
> select_features(), or have that function return an error code if strdup
> fails, and bail out early in setup_map() in that case.
> 
> Then we simply need to add a comment in assemble_map() saying that 
> mp->features must be non-null when this function is called.
> 
> As I said earlier, I'm of course not against checking function
> parameters, but here we should fail to setup a "struct multipath" in
> the first place in setup map(), rather than returning an incompletely
> initialized one. If we handle it this way, we don't need to check the
> fields of struct multipath over and over again. Similar arguments hold
> for other structs.
> 
> Of course this kind of assumption needs to be better documented in the
> code.

I'm fine with that.

-Ben

> Regards
> Martin
>