On Mon, Feb 03, 2020 at 01:55:40PM +0100, Vitaly Kuznetsov wrote:
> Sean Christopherson <sean.j.christopherson@xxxxxxxxx> writes:
>
> > diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> > index b1c469446b07..47ce04762c20 100644
> > --- a/arch/x86/kvm/cpuid.c
> > +++ b/arch/x86/kvm/cpuid.c
> > @@ -908,9 +908,14 @@ int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
> > goto out_free;
> >
> > limit = cpuid_entries[nent - 1].eax;
> > - for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
> > + for (func = ent->func + 1; func <= limit && r == 0; ++func) {
> > + if (nent >= cpuid->nent) {
> > + r = -E2BIG;
> > + goto out_free;
> > + }
> > r = do_cpuid_func(&cpuid_entries[nent], func,
> > &nent, cpuid->nent, type);
> > + }
> >
> > if (r)
> > goto out_free;
>
> Is fixing a bug a valid reason for breaking buggy userspace? :-)
> Personally, I think so.
Linus usually disagrees :-)
> In particular, here the change is both the
> return value and the fact that we don't do copy_to_user() anymore so I
> think it's possible to meet a userspace which is going to get broken by
> the change.
Ugh, yeah, it would be possible. Qemu (retries), CrosVM (hardcoded to
256 entries) and Firecracker (doesn't use the ioctl()) are all ok,
hopefully all other VMMs used in production environments follow suit.
