Re: [Qemu-devel] [PATCH] amd_iommu: Return error on machines with no PCI

[Mohammed Gamal <mgamal@redhat.com>]

On Fri, 2017-09-15 at 10:07 +0800, Peter Xu wrote:
> On Thu, Sep 14, 2017 at 05:31:38PM -0300, Eduardo Habkost wrote:
> > On Thu, Sep 14, 2017 at 10:24:23PM +0200, Thomas Huth wrote:
> > > On 14.09.2017 22:18, Mohammed Gamal wrote:
> > > > Starting the following command line causes a segfault
> > > > qemu-system-x86_64 -S -machine isapc,accel=kvm -device amd-iommu
> > > > 
> > > > This is due to the fact that the machine type 'isapc' doesn't have
> > > > a PCI bus, while amd_iommu doesn't check if the machine has PCI support
> > > > and subsequently does a null-pointer access. AMD IOMMU shouldn't even work
> > > > if the target machine doesn't have PCI.
> > > > 
> > > > Add a check for PCI on the given machine type and return an error if PCI
> > > > is not supported.
> > > > 
> > > > Signed-off-by: Mohammed Gamal <mgamal@redhat.com>
> > > > ---
> > > >  hw/i386/amd_iommu.c | 7 +++++++
> > > >  1 file changed, 7 insertions(+)
> > > > 
> > > > diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
> > > > index 334938a..9a667b7 100644
> > > > --- a/hw/i386/amd_iommu.c
> > > > +++ b/hw/i386/amd_iommu.c
> > > > @@ -1153,6 +1153,13 @@ static void amdvi_realize(DeviceState *dev, Error **err)
> > > >      }
> > > >  
> > > >      bus = pcms->bus;
> > > > +
> > > > +    if (!bus) {
> > > > +        error_setg(err, "Machine-type '%s' does not support PCI",
> > > > +                   mc->name);
> > > > +        return;
> > > > +    }
> > > > +
> > > >      s->iotlb = g_hash_table_new_full(amdvi_uint64_hash,
> > > >                                       amdvi_uint64_equal, g_free, g_free);
> > > >  
> > > > 
> > > 
> > > Patch looks fine to me, but I think it would also be sufficient to
> > > change the check at the beginning of the function to test "if (!pcms ||
> > > !pcms->bus)" instead of just "if (!pcms)" ... the error message
> > > "Machine-type 'xxx' not supported by amd-iommu" is also adequate if
> > > there is no PCI bus available on the system.
> > 
> > I agree this would be much simpler.
> 
> Even, shall we move the pcms && bus check into x86_iommu_realize()
> directly?  Then we will only need one single patch for Intel/AMD, and
> it's also a cleanup.  Thanks,

Although it's more straight forward to do the checks in amdvi_realize()
and vtd_realize() at the moment, I think moving the checks to
x86_iommu_realize() would be better on the longer term. I will be
sending out a new patch with this change shortly.

Regards,
Mohammed