On Wed, Oct 30, 2019 at 02:26:21PM -0500, Eric W. Biederman wrote:
> Andrea Righi <andrea.righi@xxxxxxxxxxxxx> writes:
>
> > On Tue, Oct 29, 2019 at 02:02:11PM -0500, Eric W. Biederman wrote:
> >> Dan Carpenter <dan.carpenter@xxxxxxxxxx> writes:
> >>
> >> > The "fix" struct has a 2 byte hole after ->ywrapstep and the
> >> > "fix = info->fix;" assignment doesn't necessarily clear it. It depends
> >> > on the compiler.
> >> >
> >> > Fixes: 1f5e31d7e55a ("fbmem: don't call copy_from/to_user() with mutex held")
> >> > Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> >> > ---
> >> > I have 13 more similar places to patch... I'm not totally sure I
> >> > understand all the issues involved.
> >>
> >> What I have done in a similar situation with struct siginfo, is that
> >> where the structure first appears I have initialized it with memset,
> >> and then field by field.
> >>
> >> Then when the structure is copied I copy the structure with memcpy.
> >>
> >> That ensures all of the bytes in the original structure are initialized
> >> and that all of the bytes are copied.
> >>
> >> The goal is to avoid memory that has values of the previous users of
> >> that memory region from leaking to userspace. Which depending on who
> >> the previous user of that memory region is could tell userspace
> >> information about what the kernel is doing that it should not be allowed
> >> to find out.
> >>
> >> I tried to trace through where "info" and thus presumably "info->fix" is
> >> coming from and only made it as far as register_framebuffer. Given
> >> that I suspect a local memset, and then a field by field copy right
> >> before copy_to_user might be a sound solution. But ick. That is a lot
> >> of fields to copy.
> >
> > I know it might sound quite inefficient, but what about making struct
> > fb_fix_screeninfo __packed?
> >
> > This doesn't solve other potential similar issues, but for this
> > particular case it could be a reasonable and simple fix.
>
> It is part of the user space ABI. As such you can't move the fields.
>
> Eric
Oh, that's right! Then memset() + memcpy() is probably the best option,
since copying all those fields one by one looks quite ugly to me...
-Andrea
