Re: Questions about commit "ipc/shm: Fix shmat mmap nil-page protection"

[Davidlohr Bueso]

On Mon, 25 Sep 2017, Andrea Arcangeli wrote:

Sorry this took so long guys. I had forgotten about this until it recently
resurfaced.

> Hello,
>
> On Mon, Sep 25, 2017 at 03:38:07PM -0400, Joe Lawrence wrote:
> > Hi Davidlohr,
> >
> > I was looking into backporting commit 95e91b831f87 ("ipc/shm: Fix shmat
> > mmap nil-page protection") to a distro kernel and Andrea brought up some
> > interesting questions about that change.
> >
> > We saw that a LTP test [1] was added some time ago to reproduce behavior
> > matching that of the original report [2].  However, Andrea and I are a
> > little confused about that original report and what the upstream commit
> > was intended to fix.  A quick summary of our offlist discussion:
> >
> > - This is only about privileged users (and no SELinux).
> >
> > - We modified the 20170119_shmat_nullpage_poc.c reproducer from [2] to
> >   include MAP_FIXED to prove (as root, no SELinux):
> >
> >     It is possible to mmap 0
> >     It is NOT possible to mmap 1
> >
> > - Andrea points out that mmap(1, ...) fails not because of any
> >   mmap_min_addr checks, but for alignment reasons.
> >
> > - He also wonders about other bogus addr values above 4k, but below
> >   mmap_min_addr and whether this change misses those values
>
> Yes, thanks for the accurate summary Joe.
>
> > Is it possible that the original report noticed that shmat allowed
> > attach to an address of 1, and it was assumed that somehow mmap_min_addr
> > protections were circumvented?  Then commit 95e91b831f87 modified the
> > rounding in do_shmat() so that shmat would fail on similar input (but
> > for apparently different reasons)?
> >
> > I didn't see any discussion when looking up the original commit in the
> > list archives, so any explanations or pointers would be very helpful.
>
> We identified only one positive side effect to such change, it is
> about the semantics of SHM_REMAP when addr < shmlba (and != 0). Before
> the patch SHM_REMAP was erroneously implicit for that virtual
> range. However that's not security related either, and there's no
> mention of SHM_REMAP in the commit message.

Coincidence. I didn't notice the SHM_REMAP, but after looking at it
you appear to be right. I'll send a patch along with the revert
(see below).

> So then we wondered what this CVE is about in the first place, it
> looks a invalid CVE for a not existent security issue. The testcase at
> least shows no malfunction, mapping addr 0 is fine to succeed with
> CAP_SYS_RAWIO.

This is exactly the issue. I thought mapping addr=0 with MAP_FIXED
was an issue, including for root. Hence avoiding the round off from
1 to 0. If this is legal, then this commit needs reverted.

In fact, X11[1] seems to rely on this _exact_ case; and this change
breaks semantics.

> From the commit message, testcase and CVE I couldn't get what this
> commit is about.
>
> Last but not the least, if there was a security problem in calling
> do_mmap_pgoff with addr=0, flags=MAP_FIXED|MAP_SHARED the fix would
> better be moved to do_mmap_pgoff, not in ipc/shm.c.

Yeah at the time, akpm and I wondered why this was special to security.

[1] https://cgit.freedesktop.org/xorg/xserver/tree/hw/xfree86/os-support/linux/int10/linux.c#n347

Thanks,
Davidlohr