Re: [PATCH v2 3/4] KVM: async_pf: Force a nested vmexit if the injected #PF is async_pf

["Radim Krčmář" <rkrcmar@redhat.com>]

2017-06-20 05:47+0800, Wanpeng Li:
> 2017-06-19 22:51 GMT+08:00 Radim Krčmář <rkrcmar@redhat.com>:
> > 2017-06-17 13:52+0800, Wanpeng Li:
> >> 2017-06-16 23:38 GMT+08:00 Radim Krčmář <rkrcmar@redhat.com>:
> >> > 2017-06-16 22:24+0800, Wanpeng Li:
> >> >> 2017-06-16 21:37 GMT+08:00 Radim Krčmář <rkrcmar@redhat.com>:
> >> >> > 2017-06-14 19:26-0700, Wanpeng Li:
> >> >> >> From: Wanpeng Li <wanpeng.li@hotmail.com>
> >> >> >>
> >> >> >> Add an async_page_fault field to vcpu->arch.exception to identify an async
> >> >> >> page fault, and constructs the expected vm-exit information fields. Force
> >> >> >> a nested VM exit from nested_vmx_check_exception() if the injected #PF
> >> >> >> is async page fault.
> >> >> >>
> >> >> >> Cc: Paolo Bonzini <pbonzini@redhat.com>
> >> >> >> Cc: Radim Krčmář <rkrcmar@redhat.com>
> >> >> >> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
> >> >> >> ---
> >> >> >> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> >> >> >> @@ -452,7 +452,11 @@ EXPORT_SYMBOL_GPL(kvm_complete_insn_gp);
> >> >> >>  void kvm_inject_page_fault(struct kvm_vcpu *vcpu, struct x86_exception *fault)
> >> >> >>  {
> >> >> >>       ++vcpu->stat.pf_guest;
> >> >> >> -     vcpu->arch.cr2 = fault->address;
> >> >> >> +     vcpu->arch.exception.async_page_fault = fault->async_page_fault;
> >> >> >
> >> >> > I think we need to act as if arch.exception.async_page_fault was not
> >> >> > pending in kvm_vcpu_ioctl_x86_get_vcpu_events().  Otherwise, if we
> >> >> > migrate with pending async_page_fault exception, we'd inject it as a
> >> >> > normal #PF, which could confuse/kill the nested guest.
> >> >> >
> >> >> > And kvm_vcpu_ioctl_x86_set_vcpu_events() should clean the flag for
> >> >> > sanity as well.
> >> >>
> >> >> Do you mean we should add a field like async_page_fault to
> >> >> kvm_vcpu_events::exception, then saves arch.exception.async_page_fault
> >> >> to events->exception.async_page_fault through KVM_GET_VCPU_EVENTS and
> >> >> restores events->exception.async_page_fault to
> >> >> arch.exception.async_page_fault through KVM_SET_VCPU_EVENTS?
> >> >
> >> > No, I thought we could get away with a disgusting hack of hiding the
> >> > exception from userspace, which would work for migration, but not if
> >> > local userspace did KVM_GET_VCPU_EVENTS and KVM_SET_VCPU_EVENTS ...
> >> >
> >> > Extending the userspace interface would work, but I'd do it as a last
> >> > resort, after all conservative solutions have failed.
> >> > async_pf migration is very crude, so exposing the exception is just an
> >> > ugly workaround for the local case.  Adding the flag would also require
> >> > userspace configuration of async_pf features for the guest to keep
> >> > compatibility.
> >> >
> >> > I see two options that might be simpler than adding the userspace flag:
> >> >
> >> >  1) do the nested VM exit sooner, at the place where we now queue #PF,
> >> >  2) queue the #PF later, save the async_pf in some intermediate
> >> >     structure and consume it at the place where you proposed the nested
> >> >     VM exit.
> >>
> >> How about something like this to not get exception events if it is
> >> "is_guest_mode(vcpu) && vcpu->arch.exception.nr == PF_VECTOR &&
> >> vcpu->arch.exception.async_page_fault" since lost a reschedule
> >> optimization is not that importmant in L1.
> >>
> >> @@ -3072,13 +3074,16 @@ static void
> >> kvm_vcpu_ioctl_x86_get_vcpu_events(struct kvm_vcpu *vcpu,
> >>                             struct kvm_vcpu_events *events)
> >>  {
> >>      process_nmi(vcpu);
> >> -    events->exception.injected =
> >> -        vcpu->arch.exception.pending &&
> >> -        !kvm_exception_is_soft(vcpu->arch.exception.nr);
> >> -    events->exception.nr = vcpu->arch.exception.nr;
> >> -    events->exception.has_error_code = vcpu->arch.exception.has_error_code;
> >> -    events->exception.pad = 0;
> >> -    events->exception.error_code = vcpu->arch.exception.error_code;
> >> +    if (!(is_guest_mode(vcpu) && vcpu->arch.exception.nr == PF_VECTOR &&
> >> +        vcpu->arch.exception.async_page_fault)) {
> >> +        events->exception.injected =
> >> +            vcpu->arch.exception.pending &&
> >> +            !kvm_exception_is_soft(vcpu->arch.exception.nr);
> >> +        events->exception.nr = vcpu->arch.exception.nr;
> >> +        events->exception.has_error_code = vcpu->arch.exception.has_error_code;
> >> +        events->exception.pad = 0;
> >> +        events->exception.error_code = vcpu->arch.exception.error_code;
> >> +    }
> >
> > This adds a bug when userspace does KVM_GET_VCPU_EVENTS and
> > KVM_SET_VCPU_EVENTS without migration -- KVM would drop the async_pf and
> > a L1 process gets stuck as a result.
> >
> > We we'd need to add a similar condition to
> > kvm_vcpu_ioctl_x86_set_vcpu_events(), so userspace SET doesn't drop it,
> > but that is far beyond the realm of acceptable code.
> 
> Do you mean current status of the patchset v2 can be accepted?
> Otherwise, what's the next should be done?

No, sorry, that one has the migration bug (the async_page_fault gets
dropped on destination).

You proposed to add the flag to the userspace interface, which is a
sound solution.  I was asking to look for a different one, because the
flag is a work-around for an implementation detail, which isn't a good
thing to put into a userspace interface ...

Still, I looked at the early VM exit (1) and it doesn't fit well into
SVM's model of single nested VM exit location, so it's out.

The remaining contender is to add a paravirtualized event for apf and
only convert it into nested VM exit or #PF in inject_pending_event().
The end result would likely be a slightly better version of the
exception flag ...

I guess that doing a prototype of the userspace interface extension is a
good follow up.