On Fri, Mar 05, 2021, Ashish Kalra wrote:
> On Thu, Feb 25, 2021 at 10:49:00AM -0800, Steve Rutherford wrote:
> > On Thu, Feb 25, 2021 at 6:57 AM Tom Lendacky <thomas.lendacky@xxxxxxx> wrote:
> > > >> +int svm_vm_copy_asid_to(struct kvm *kvm, unsigned int mirror_kvm_fd)
> > > >> +{
> > > >> + struct file *mirror_kvm_file;
> > > >> + struct kvm *mirror_kvm;
> > > >> + struct kvm_sev_info *mirror_kvm_sev;
> > > >> + unsigned int asid;
> > > >> + int ret;
> > > >> +
> > > >> + if (!sev_guest(kvm))
> > > >> + return -ENOTTY;
> > > >
> > > > You definitely don't want this: this is the function that turns the vm
> > > > into an SEV guest (marks SEV as active).
> > >
> > > The sev_guest() function does not set sev->active, it only checks it. The
> > > sev_guest_init() function is where sev->active is set.
> > Sorry, bad use of the english on my part: the "this" was referring to
> > svm_vm_copy_asid_to. Right now, you could only pass this sev_guest
> > check if you had already called sev_guest_init, which seems incorrect.
> > >
> > > >
> > > > (Not an issue with this patch, but a broader issue) I believe
> > > > sev_guest lacks the necessary acquire/release barriers on sev->active,
> > >
> > > The svm_mem_enc_op() takes the kvm lock and that is the only way into the
> > > sev_guest_init() function where sev->active is set.
> > There are a few places that check sev->active which don't have the kvm
> > lock, which is not problematic if we add in a few compiler barriers
> > (ala irqchip_split et al).
Eh, I don't see the point in taking on the complexity of barriers. Ignoring the
vCPU behavior, the only existing call that isn't safe is svm_register_enc_region().
Fixing that is trivial and easy to understand.
As for the vCPU stuff, adding barriers will not make them safe. E.g. a barrier
won't magically make init_vmcb() go back in time and set SVM_NESTED_CTL_SEV_ENABLE
if SEV is enabled after vCPUs are created.
> Probably, sev->active accesses can be made safe using READ_ONCE() &
> WRITE_ONCE().
