[PATCH v3] fs/fcntl: restore checking against COMPAT_LOFF_T_MAX for F_GETLK64

[Jeff Layton]

On Wed, 2017-11-15 at 00:22 +0300, Vitaly Lipatov wrote:
> Jeff Layton писал 14.11.17 23:19:
> > On Tue, 2017-11-14 at 22:25 +0300, Vitaly Lipatov wrote:
> > > Jeff Layton писал 14.11.17 22:12:
> > > ...
> > > > Wait...
> > > > 
> > > > Does this do anything at all in the case where you pass in
> > > > COMPAT_LOFF_T_MAX? l_start and l_len are either off_t or loff_t
> > > > (depending on arch).
> > > > 
> > > > Either one will fit in the F_GETLK64/F_OFD_GETLK struct, so I don't see
> > > > a need to check here.
> > > 
> > > I am not sure, can off_t be bigger than loff_t ?
> > 
> > I don't think so, at least not in any possible situation we care about
> > here.
> 
> We have this checking for ages:
> 			if (f.l_start > COMPAT_LOFF_T_MAX)
>   				ret = -EOVERFLOW;
> http://debian.securedservers.com/kernel/pub/linux/kernel/people/akpm/patches/2.6/2.6.15-rc5/2.6.15-rc5-mm1/broken-out/fix-overflow-tests-for-compat_sys_fcntl64-locking.patch
> 

I'm not convinced that those checks ever did anything, tbh.

> > 
> > > If not, we have just skip checking against COMPAT_LOFF_T_MAX.
> > > 
> > > ...
> > > > > @@ -644,7 +644,7 @@ COMPAT_SYSCALL_DEFINE3(fcntl64, unsigned int, fd,
> > > > > unsigned int, cmd,
> > > > >  		err = fcntl_getlk(f.file, convert_fcntl_cmd(cmd), &flock);
> > > > >  		if (err)
> > > > >  			break;
> > > > > -		err = fixup_compat_flock(&flock);
> > > > > +		err = fixup_compat_flock(&flock, COMPAT_LOFF_T_MAX);
> > > > >  		if (err)
> > > > >  			return err;
> > > > >  		err = put_compat_flock64(&flock, compat_ptr(arg));
> > > > 
> > > > Maybe a simpler fix would be to just remove the fixup_compat_flock call
> > > > above?
> > > > 
> > 
> > Ok. If you have a test for this, mind testing and sending a patch?
> 
> I think if COMPAT_LOFF_T_MAX is exists, that value can be smaller than 
> can fit in off_t.
> Checking against COMPAT_LOFF_T_MAX keep old logic works for me last 10 
> years.
> 
> I have some tests around wine project I worked on. May be later I will 
> do additional tests.
> 

I am making an assumption here that l_start and l_end can never be
larger than a signed 64-bit value. I don't see how it ever could be,
given that it's defined as a long long, but I suppose we could add some
exotic arch later that does something weird.

Maybe we can just add a BUILD_BUG_ON for that? I'll send along an
alternate patch in a few mins.
-- 
Jeff Layton <jlayton at kernel.org>