Re: strict copy_from_user checks issues?

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 05 January 2010, Heiko Carstens wrote:
> On Mon, Jan 04, 2010 at 05:43:08PM -0800, Arjan van de Ven wrote:
> > On Mon, 4 Jan 2010 16:43:45 +0100
> > Heiko Carstens <heiko.carstens@de.ibm.com> wrote:
> > > x86 and sparc return -EFAULT in copy_from_user instead of the number
> > > of not copied bytes as it should in case of a detected buffer
> > > overflow. That might have unwanted side effects. I would guess that
> > > is a bug.
> > 
> > killing the bad guy in case of a real buffer overflow is appropriate..
> > this should never trigger for legitimate users.
> 
> The point I tried to make is that no caller of copy_from_user can assume
> that it would ever return -EFAULT. And if any caller does so it is broken.
> But then again it probably doesn't matter in this case as long as something
> != 0 is returned.

To quote simple_read_from_buffer():

        size_t ret;
	...
        ret = copy_to_user(to, from + pos, count);
        if (ret == count)
                return -EFAULT;
        count -= ret;
        *ppos = pos + count;
        return count;

If copy_from_user() returns a negative value, bad things happen to f_pos
and to the value returned from the syscall. Many read() file_operations
do this similarly, and I wouldn't be surprised if this could be turned
into a security exploit for one of them (not simple_read_from_buffer
probably).

	Arnd