Re: [PATCH V2 for-5.2] hw/null-machine: Add the kvm_type() hook for MIPS

[Huacai Chen]

Hi, Thomas,

On Wed, Sep 9, 2020 at 3:20 PM Thomas Huth <thuth@redhat.com> wrote:
>
> On 09/09/2020 04.57, chen huacai wrote:
> > Hi, all,
> >
> > On Wed, Sep 9, 2020 at 1:25 AM Thomas Huth <thuth@redhat.com> wrote:
> >>
> >> On 24/08/2020 10.11, Huacai Chen wrote:
> >>> MIPS has two types of KVM: TE & VZ, and TE is the default type. Now,
> >>> libvirt uses a null-machine to detect the kvm capability. In the MIPS
> >>> case, it will return "KVM not supported" on a VZ platform by default.
> >>> So, add the kvm_type() hook to the null-machine.
> >>>
> >>> This seems not a very good solution, but I cannot do it better now.
> >>
> >> This is still ugly. Why do the other architectures do not have the
> >> same problem? Let's see... in kvm-all.c, we have:
> >>
> >>     int type = 0;
> >>     [...]
> >>     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
> >>     if (mc->kvm_type) {
> >>         type = mc->kvm_type(ms, kvm_type);
> >>     } else if (kvm_type) {
> >>         ret = -EINVAL;
> >>         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
> >>         goto err;
> >>     }
> >>
> >>     do {
> >>         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
> >>     } while (ret == -EINTR);
> >>
> >> Thus the KVM_CREATE_VM ioctl is likely called with type = 0 in this
> >> case (i.e. when libvirt probes with the "null"-machine).
> >>
> >> Now let's have a look at the kernel. The "type" parameter is passed
> >> there to the architecture specific function kvm_arch_init_vm().
> >> For powerpc, this looks like:
> >>
> >>         if (type == 0) {
> >>                 if (kvmppc_hv_ops)
> >>                         kvm_ops = kvmppc_hv_ops;
> >>                 else
> >>                         kvm_ops = kvmppc_pr_ops;
> >>                 if (!kvm_ops)
> >>                         goto err_out;
> >>         } else  if (type == KVM_VM_PPC_HV) {
> >>                 if (!kvmppc_hv_ops)
> >>                         goto err_out;
> >>                 kvm_ops = kvmppc_hv_ops;
> >>         } else if (type == KVM_VM_PPC_PR) {
> >>                 if (!kvmppc_pr_ops)
> >>                         goto err_out;
> >>                 kvm_ops = kvmppc_pr_ops;
> >>         } else
> >>                 goto err_out;
> >>
> >> That means for type == 0, it automatically detects the best
> >> kvm-type.
> >>
> >> For mips, this function looks like this:
> >>
> >>         switch (type) {
> >> #ifdef CONFIG_KVM_MIPS_VZ
> >>         case KVM_VM_MIPS_VZ:
> >> #else
> >>         case KVM_VM_MIPS_TE:
> >> #endif
> >>                 break;
> >>         default:
> >>                 /* Unsupported KVM type */
> >>                 return -EINVAL;
> >>         };
> >>
> >> That means, for type == 0, it returns -EINVAL here!
> >>
> >> Looking at the API docu in Documentation/virt/kvm/api.rst
> >> the description of the type parameter is quite sparse, but it
> >> says:
> >>
> >>  "You probably want to use 0 as machine type."
> >>
> >> So I think this is a bug in the implementation of KVM in the
> >> mips kernel code. The kvm_arch_init_vm() in the mips code should
> >> do the same as on powerpc, and use the best available KVM type
> >> there instead of returning EINVAL. Once that is fixed there,
> >> you don't need this patch here for QEMU anymore.
> > Yes, PPC use a good method, because it can use 0 as "automatic"
> > #define KVM_VM_PPC_HV 1
> > #define KVM_VM_PPC_PR 2
> >
> > Unfortunately, MIPS cannot do like this because it define 0 as "TE":
> > #define KVM_VM_MIPS_TE          0
> > #define KVM_VM_MIPS_VZ          1
> >
> > So, it cannot be solved in kernel side, unless changing the definition
> > of TE/VZ, but I think changing their definition is also unacceptable.
>
> Ouch, ok, now I understood the problem. That sounds like a really bad
> decision on the kernel side.
>
> But I think you could at least try to get it fixed on the kernel side:
> Propose a patch to rename KVM_VM_MIPS_TE to KVM_VM_MIPS_DEFAULT and use
> 2 as new value for TE. The code that handles the default 0 case should
> then prefer TE over VZ, so that old userspace applications that provide
> 0 will continue to work as they did before, so I hope that the change is
> acceptable by the kernel folks. You should add a remark to the patch
> description that 0 is the established default value for probing in the
> QEMU/libvirt stack and that your patch is required on the kernel side to
> avoid ugly hacks in QEMU userspace code.
>
> If they still reject your patch, I guess we have to bite the bullet and
> add your patch here...
OK, let me try.

Huacai
>
> Thanks,
>  Thomas
>