Re: [RFC PATCH-for-7.2 4/5] hw/display/qxl: Avoid buffer overrun in qxl_phys2virt (CVE-2022-4144)

[Philippe Mathieu-Daudé]

On 28/11/22 16:32, Stefan Hajnoczi wrote:
> On Mon, 28 Nov 2022 at 10:25, Philippe Mathieu-Daudé <philmd@linaro.org> wrote:
> > On 28/11/22 16:16, Stefan Hajnoczi wrote:
> > > On Mon, 28 Nov 2022 at 08:53, Philippe Mathieu-Daudé <philmd@linaro.org> wrote:
> > > > Have qxl_get_check_slot_offset() return false if the requested
> > > > buffer size does not fit within the slot memory region.
> > > >
> > > > Similarly qxl_phys2virt() now returns NULL in such case, and
> > > > qxl_dirty_one_surface() aborts.
> > > >
> > > > This avoids buffer overrun in the host pointer returned by
> > > > memory_region_get_ram_ptr().
> > > >
> > > > Fixes: CVE-2022-4144 (out-of-bounds read)
> > > > Reported-by: Wenxu Yin (@awxylitol)
> > > > Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1336
> > > > Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> > > > ---
> > > >    hw/display/qxl.c | 22 ++++++++++++++++++----
> > > >    hw/display/qxl.h |  2 +-
> > > >    2 files changed, 19 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/hw/display/qxl.c b/hw/display/qxl.c
> > > > index 231d733250..afa157d327 100644
> > > > --- a/hw/display/qxl.c
> > > > +++ b/hw/display/qxl.c
> > > > @@ -1424,11 +1424,13 @@ static void qxl_reset_surfaces(PCIQXLDevice *d)
> > > >
> > > >    /* can be also called from spice server thread context */
> > > >    static bool qxl_get_check_slot_offset(PCIQXLDevice *qxl, QXLPHYSICAL pqxl,
> > > > -                                      uint32_t *s, uint64_t *o)
> > > > +                                      uint32_t *s, uint64_t *o,
> > > > +                                      size_t size_requested)
> > > >    {
> > > >        uint64_t phys   = le64_to_cpu(pqxl);
> > > >        uint32_t slot   = (phys >> (64 -  8)) & 0xff;
> > > >        uint64_t offset = phys & 0xffffffffffff;
> > > > +    uint64_t size_available;
> > > >
> > > >        if (slot >= NUM_MEMSLOTS) {
> > > >            qxl_set_guest_bug(qxl, "slot too large %d >= %d", slot,
> > > > @@ -1453,6 +1455,18 @@ static bool qxl_get_check_slot_offset(PCIQXLDevice *qxl, 
> > > > QXLPHYSICAL pqxl,
> > > >            return false;
> > > >        }
> > > >
> > > > +    size_available = memory_region_size(qxl->guest_slots[slot].mr);
> > > > +    assert(qxl->guest_slots[slot].offset + offset < size_available);
> > >
> > > Can this assertion be triggered by the guest (via an invalid pqxl
> > > value)? I think the answer is no, but I don't know the the qxl code
> > > well enough to be sure.
> >
> > 'qxl->guest_slots[slot].offset' is initialized in qxl_add_memslot()
> > (host); 'size_available' also comes from the host, but 'offset'
> > comes from the guest via 'QXLPHYSICAL pqxl' IIUC.
> >
> > I added this check to avoid overflow, but it can be changed to return
> > an error.
>
> Yes, please.

Or I could use Int128 to do arithmetic, but various other places do it
this way without checking overflow with memory_region_size(). Such API
change should be global and is out of the scope of this CVE fix IMO.

> Aside from concerns about -DNDEBUG, which builds without assertions,

This isn't an issue anymore since 262a69f428 ("osdep.h: Prohibit 
disabling assert() in supported builds").

> there is also a DoS issue with nested virt where an L2 guest shouldn't
> be able to abort the L1 guest's QEMU by triggering an assertion in a
> pass through device.
>
> Guest input validation should use explicit error checking code instead
> of assert(3).

Certainly.