Re: [PATCH 1/5] KVM: x86: Get vmcs12 pages before checking pending interrupts

[Sean Christopherson <seanjc@google.com>]

On Thu, Aug 25, 2022, Mingwei Zhang wrote:
> On Thu, Aug 25, 2022 at 7:41 AM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Wed, Aug 24, 2022, Jim Mattson wrote:
> > > On Wed, Aug 24, 2022 at 5:11 PM Sean Christopherson <seanjc@google.com> wrote:
> > >
> > > > @google folks, what would it take for us to mark KVM_REQ_GET_NESTED_STATE_PAGES
> > > > as deprecated in upstream and stop accepting patches/fixes?  IIUC, when we eventually
> > > > move to userfaultfd, all this goes away, i.e. we do want to ditch this at some point.
> > >
> > > Userfaultfd is a red herring. There were two reasons that we needed
> > > this when nested live migration was implemented:
> > > 1) our netlink socket mechanism for funneling remote page requests to
> > > a userspace listener was broken.
> > > 2) we were not necessarily prepared to deal with remote page requests
> > > during VM setup.
> > >
> > > (1) has long since been fixed. Though our preference is to exit from
> > > KVM_RUN and get the vCPU thread to request the remote page itself, we
> > > are now capable of queuing a remote page request with a separate
> > > listener thread and blocking in the kernel until the page is received.
> > > I believe that mechanism is functionally equivalent to userfaultfd,
> > > though not as elegant.
> > > I don't know about (2). I'm not sure when the listener thread is set
> > > up, relative to all of the other setup steps. Eliminating
> > > KVM_REQ_GET_NESTED_STATE_PAGES means that userspace must be prepared
> > > to fetch a remote page by the first call to KVM_SET_NESTED_STATE. The
> > > same is true when using userfaultfd.
> > >
> > > These new ordering constraints represent a UAPI breakage, but we don't
> > > seem to be as concerned about that as we once were. Maybe that's a
> > > good thing. Can we get rid of all of the superseded ioctls, like
> > > KVM_SET_CPUID, while we're at it?
> >
> > I view KVM_REQ_GET_NESTED_STATE_PAGES as a special case.  We are likely the only
> > users, we can (eventually) wean ourselves off the feature, and we can carry
> > internal patches (which we are obviously already carrying) until we transition
> > away.  And unlike KVM_SET_CPUID and other ancient ioctls() that are largely
> > forgotten, this feature is likely to be a maintenance burden as long as it exists.
> 
> KVM_REQ_GET_NESTED_STATE_PAGES has been uniformly used in
> KVM_SET_NESTED_STATE ioctl in VMX (including eVMCS) and SVM, it is
> basically a two-step setting up of a nested state mechanism.
> 
> We can change that, but this may have side effects and I think this
> usage case has nothing to do with demand paging.

There are two uses of KVM_REQ_GET_NESTED_STATE_PAGES:

  1. Defer loads when leaving SMM.

  2: Defer loads for KVM_SET_NESTED_STATE.

#1 is fully solvable without a request, e.g. split ->leave_smm() into two helpers,
one that restores whatever metadata is needed before restoring from SMRAM, and
a second to load guest virtualization state that _after_ restoring all other guest
state from SMRAM.

#2 is done because of the reasons Jim listed above, which are specific to demand
paging (including userfaultfd).  There might be some interactions with other
ioctls() (KVM_SET_SREGS?) that are papered over by the request, but that can be
solved without a full request since only the first KVM_RUN after KVM_SET_NESTED_STATE
needs to refresh things (though ideally we'd avoid that).

In other words, if the demand paging use case goes away, then KVM can get rid of
KVM_REQ_GET_NESTED_STATE_PAGES.  

> KVM_SET_NESTED_STATE in VMX, while in SVM implementation, it is simply
> just a kvm_make_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);

svm_set_nested_state() very rougly open codes enter_svm_guest_mode().  VMX could
do the same, but that may or may not be a net positive.

> hmm... so is the nested_vmx_enter_non_root_mode() call in vmx
> KVM_SET_NESTED_STATE ioctl() still necessary? I am thinking that
> because the same function is called again in nested_vmx_run().

nested_vmx_run() is used only to emulate VMLAUNCH/VMRESUME and wont' be invoked
if the vCPU is already running L2 at the time of migration.