On Tue, Nov 12, 2019 at 06:45:55AM -0800, Ram Pai wrote: > On Tue, Nov 12, 2019 at 10:32:04PM +1100, Paul Mackerras wrote: > > On Mon, Nov 11, 2019 at 11:52:15PM -0800, Ram Pai wrote: > > > There is subtle problem removing that code from the assembly. > > > > > > If the H_SVM_INIT_ABORT hcall returns to the ultravisor without clearing > > > kvm->arch.secure_guest, the hypervisor will continue to think that the > > > VM is a secure VM. However the primary reason the H_SVM_INIT_ABORT > > > hcall was invoked, was to inform the Hypervisor that it should no longer > > > consider the VM as a Secure VM. So there is a inconsistency there. > > > > Most of the checks that look at whether a VM is a secure VM use code > > like "if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)". Now > > since KVMPPC_SECURE_INIT_ABORT is 4, an if statement such as that will > > take the false branch once we have set kvm->arch.secure_guest to > > KVMPPC_SECURE_INIT_ABORT in kvmppc_h_svm_init_abort. So in fact in > > most places we will treat the VM as a normal VM from then on. If > > there are any places where we still need to treat the VM as a secure > > VM while we are processing the abort we can easily do that too. > > Is the suggestion -- KVMPPC_SECURE_INIT_ABORT should never return back > to the Ultravisor? Because removing that assembly code will NOT lead the No. The suggestion is that vcpu->arch.secure_guest stays set to KVMPPC_SECURE_INIT_ABORT until userspace calls KVM_PPC_SVM_OFF. > Hypervisor back into the Ultravisor. This is fine with the > ultravisor. But then the hypervisor will not know where to return to. > If it wants to return directly to the VM, it wont know to > which address. It will be in a limbo. > > > > > > This is fine, as long as the VM does not invoke any hcall or does not > > > receive any hypervisor-exceptions. The moment either of those happen, > > > the control goes into the hypervisor, the hypervisor services > > > the exception/hcall and while returning, it will see that the > > > kvm->arch.secure_guest flag is set and **incorrectly** return > > > to the ultravisor through a UV_RETURN ucall. Ultravisor will > > > not know what to do with it, because it does not consider that > > > VM as a Secure VM. Bad things happen. > > > > If bad things happen in the ultravisor because the hypervisor did > > something it shouldn't, then it's game over, you just lost, thanks for > > playing. The ultravisor MUST be able to cope with bogus UV_RETURN > > calls for a VM that it doesn't consider to be a secure VM. You need > > to work out how to handle such calls safely and implement that in the > > ultravisor. > > Actually we do handle this gracefully in the ultravisor :). > We just retun back to the hypervisor saying "sorry dont know what > to do with it, please handle it yourself". > > However hypervisor would not know what to do with that return, and bad > things happen in the hypervisor. Right. We need something after the "sc 2" to handle the case where the ultravisor returns with an error from the UV_RETURN. Paul.
