On Mon, Oct 17, 2022, Peter Gonda wrote: > On Mon, Oct 17, 2022 at 12:04 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > > On Mon, Oct 17, 2022, Peter Gonda wrote: > > > This refactor sounds good, working on this with a few changes. > > > > > > Instead of kvm_init_vm_address_properties() as you suggested I've added this: > > > > > > @@ -272,6 +275,8 @@ struct kvm_vm *____vm_create(enum vm_guest_mode > > > mode, uint64_t nr_pages) > > > vm->type = KVM_VM_TYPE_ARM_IPA_SIZE(vm->pa_bits); > > > #endif > > > > > > + kvm_init_vm_arch(vm); > > > > Why? I'm not necessarily opposed to adding kvm_init_vm_arch(), but since x86 > > "needs" a dedicated hook to unpack the mode, why not piggyback that one? > > > > Well I since I need to do more than just > kvm_init_vm_address_properties() I thought the more generic name would > be better. We need to allocate kvm_vm_arch, find the c-bit, and call > KVM_SEV_INIT. I can put it back in that switch case if thats better, > thoughts? > > > > + > > > vm_open(vm); > > > > > > /* Limit to VA-bit canonical virtual addresses. */ > > > > > > And I need to put kvm_arch_vm_post_create() after the vCPUs are > > > created because the ordering we need is: KVM_SEV_INIT -> Create vCPUS > > > -> KVM_SEV_LAUNCH_FINISH. > > > > Hrm, that's annoying. Please don't use kvm_arch_vm_post_create() as the name, > > that's a better fit for what Vishal is doing since the "vm_post_create()" implies > > that it's called for "all" VM creation paths, where "all" means "everything > > except barebones VMs". E.g. in Vishal's series, kvm_arch_vm_post_create() can > > be used to drop the vm_create_irqchip() call in common code. In your case, IIUC > > the hook will be invoked from __vm_create_with_vcpus(). > > > > I'm a little hesitant to have an arch hook for this case since it can't be > > all-or-nothing (again, ignoring barebones VMs). If a "finalize" arch hook is added, > > then arguably tests that do __vm_create() and manually add vCPUs should call the > > arch hook, i.e. we'd be adding maintenance burden to tests that in all likelihood > > don't care about SEV and never will. > > > > It's somewhat unfortunate, but dedicated vm_sev_create_with_one_vcpu() and > > and vm_sev_create_with_vcpus() wrappers is probably the least awful solution. > > Make sense. I think we can go back to your suggestion of > kvm_init_vm_address_properties() above since we can now do all the > KVM_SEV_* stuff. I think this means we don't need to add > VM_MODE_PXXV48_4K_SEV since we can set up the c-bit from inside of > vm_sev_create_*(), thoughts? Configuring the C-bit inside vm_sev_create_*() won't work (at least not well). The C-bit needs to be known before kvm_vm_elf_load(), i.e. can't be handled after __vm_create(), and needs to be tracked inside the VM, i.e. can't be handled before __vm_create(). The proposed kvm_init_vm_address_properties() seems like the best fit since the C-bit (and TDX's S-bit) is stolen from GPA space, i.e. directly affects the other values computed in that path. As for the kvm_vm_arch allocation ugliness, when we talked off-list I didn't consider the need to allocate in kvm_init_vm_address_properties(). That's quite gross, especially since the pointer will be larger than the thing being allocated. With that in mind, adding .../include/<arch>/kvm_util.h so that "struct kvm_vm_arch" can be defined and referenced directly doesn't seem so bad. Having to stub in the struct for the other architectures is annoying, but not the end of the world.
