[PATCH] fs: squasfs: fix a possible NULL pointer dereference in sqfs_opendir()

[Miquel Raynal]

Hi Richard,

Richard Genoud <richard.genoud at posteo.net> wrote on Mon, 21 Dec 2020
17:17:56 +0100:

> Hi Miquel
> 
> Le 21/12/2020 à 16:49, Miquel Raynal a écrit :
> > Hi Richard,
> > 
> > Richard Genoud <richard.genoud at posteo.net> wrote on Mon, 21 Dec 2020
> > 16:40:51 +0100:
> >   
> >> Hi Miquel,
> >>
> >> Le 21/12/2020 à 16:29, Miquel Raynal a écrit :  
> >>> Hi Richard,
> >>>
> >>> Richard Genoud <richard.genoud at posteo.net> wrote on Mon, 21 Dec 2020
> >>> 16:26:00 +0100:  
> >>>    >>>> Le 21/12/2020 à 16:14, Miquel Raynal a écrit :  
> >>>>> Hi Richard,
> >>>>>
> >>>>> Richard Genoud <richard.genoud at posteo.net> wrote on Mon, 21 Dec 2020
> >>>>> 16:06:37 +0100:  
> >>>>>     >>>> Hi Miquel,  
> >>>>>>
> >>>>>> Le 18/12/2020 à 19:50, Miquel Raynal a écrit :  
> >>>>>>> Hi Richard,
> >>>>>>>
> >>>>>>> Richard Genoud <richard.genoud at posteo.net> wrote on Fri, 18 Dec 2020
> >>>>>>> 15:24:40 +0100:  
> >>>>>>>      >>>> token_count may be != 0 and token_list not yet allocated when the out  
> >>>>>>>> code is reached  
> >>>>>>>
> >>>>>>> Wouldn't it be better to initialize token_count than adding an
> >>>>>>> (obscure) indentation level?  
> >>>>>> Well, token_count is initialized :
> >>>>>> token_count = sqfs_count_tokens(filename);
> >>>>>>
> >>>>>> But token_list is not yet populated. It is some lines bellow:
> >>>>>> token_list = malloc(token_count * sizeof(char *));
> >>>>>>
> >>>>>>
> >>>>>> But I could use something like that, maybe it's clearer :
> >>>>>> 	for (j = 0; (token_list != NULL) && (j < token_count); j++)
> >>>>>> 		free(token_list[j]);  
> >>>>>
> >>>>> I had a look at the code, the error path is clearly not correctly
> >>>>> organized.
> >>>>>
> >>>>> I think the right approach would be to have real labels like,
> >>>>> free_token_list, free_this, free_that and for each of them only do the
> >>>>> right cleanup. Doing so should fix the issue.  
> >>>>
> >>>> So you're suggesting to revert this ?
> >>>> commit ea1b1651c6a8 ("fs/squashfs: sqfs_opendir: simplify error handling")  
> >>>
> >>> Yes (our e-mails crossed each other), I think it's best to have a well
> >>> organized error path. Of course this error path is maybe faulty, in
> >>> this case it must be fixed. But I personally prefer the revert + fix
> >>> approach.  
> >>>    >>  
> >> But I really don't see why it's obscure to test a pointer before dereference.  
> > 
> > Testing a pointer before dereference is not obscure.
> > 
> > Testing a pointer in an error path because the error path is common to
> > all 10 different possible failure cases and might free the content of an
> > array that has not been allocated yet: this is obscure.
> >   
> >> Maybe I should I've wrote :
> >>          if (token_list != NULL)
> >>                  for (j = 0; j < token_count; j++)
> >>                          free(token_list[j]);
> >>
> >> Does it looks better ?  
> > 
> > Not really :)  
> 
> Ok, so if you insist, I send the revert correcting the coverity issue.
> 
> But in this case, the error management won't be coherent with the rest of the file.
> (And I *really* don't want to revert to the old error handling for every single function.)

Well, I was against taking this direction from the beginning, now we
are at a point where the error path must be fixed because you need to
take extra precautions that you would have avoided by keeping the well
organized labels.

Thanks,
Miquèl