Re: [PART1 RFC v3 02/12] KVM: x86: Introducing kvm_x86_ops VM init/uninit hooks

[Suravee Suthikulpanit]

Hi Paolo,

On 3/30/16 19:07, Paolo Bonzini wrote:
> On 30/03/2016 12:00, Suravee Suthikulpanit wrote:
> > Hi Paolo,
> >
> > On 3/29/16 18:47, Suravee Suthikulpanit wrote:
> > > Hi,
> > >
> > > On 03/29/2016 05:21 PM, Paolo Bonzini wrote:
> > > > On 29/03/2016 07:27, Suravee Suthikulpanit wrote:
> > > > > > > > Adding function pointers in struct kvm_x86_ops for
> > > > > > > > processor-specific
> > > > > > > > layer to provide hooks for when KVM initialize and un-initialize VM.
> > > > > > This is not the only thing your patch is doing, and the "other" change
> > > > > > definitely needs a lot more explanation about why you did it and how
> > > > > > you
> > > > > > audited the code to ensure that it is safe.
> > > > > >
> > > > > > Paolo
> > > > >
> > > > > Sorry, for not mentioning this earlier. I am moving the
> > > > > kvm_arch_init_vm() call mainly to go after mutex_init(&kvm->slots_lock)
> > > > > since I am calling the x86_set_memory_region() (which uses slots_lock)
> > > > > in the vm_init() hooks (for AVIC initialization).
> > > > >
> > > > > Lemme re-check if this would be safe for other code.
> > > >
> > > > No problem.  In the meanwhile a patch got in ("KVM: fix spin_lock_init
> > > > order on x86") that should help you.
> > > >
> > > > Thanks,
> > > >
> > > > Paolo
> > >
> > > Right.... that's just what I need :) I'll re-base to the latest tip then.
> >
> > Actually, in the file virt/kvm/kvm_main.c, I still need to move the
> > kvm_arch_init_vm() to some place after the call to kvm_alloc_memslots()
> > since I am calling x86_set_memory_region() in the vm_init hook.
> >
> >          r = -ENOMEM;
> >          for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
> >                  kvm->memslots[i] = kvm_alloc_memslots();
> >                  if (!kvm->memslots[i])
> >                          goto out_err_no_srcu;
> >          }
> >
> >          if (init_srcu_struct(&kvm->srcu))
> >                  goto out_err_no_srcu;
> >          if (init_srcu_struct(&kvm->irq_srcu))
> >                  goto out_err_no_irq_srcu;
> >          for (i = 0; i < KVM_NR_BUSES; i++) {
> >                  kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
> >                                          GFP_KERNEL);
> >                  if (!kvm->buses[i])
> >                          goto out_err;
> >          }
> > //HERE
> >          r = kvm_arch_init_vm(kvm, type);
> >          if (r)
> >                  goto out_err;
> >
> > Do you think that would be a problem?
>
> Can you delay that to after the creation of the first VCPU?

Sure, I can. That's what I was doing originally before we introduce the 
vm_init hook. I just thought that this would make a nice place.

> Allocating AVIC data structures is not required if userspace has not
> called KVM_CREATE_IRQCHIP or enabled KVM_CAP_SPLIT_IRQCHIP.
>
> Paolo

Okay.

Thanks,
Suravee