On Sun, Oct 10, 2010 at 05:55:25PM +0200, Avi Kivity wrote:
> >>
> >> >> >@@ -5112,6 +5122,13 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> >> >> > if (unlikely(r))
> >> >> > goto out;
> >> >> >
> >> >> >+ kvm_check_async_pf_completion(vcpu);
> >> >> >+ if (vcpu->arch.mp_state == KVM_MP_STATE_HALTED) {
> >> >> >+ /* Page is swapped out. Do synthetic halt */
> >> >> >+ r = 1;
> >> >> >+ goto out;
> >> >> >+ }
> >> >> >+
> >> >>
> >> >> Why do it here in the fast path? Can't you halt the cpu when
> >> >> starting the page fault?
> >> >Page fault may complete before guest re-entry. We do not want to halt vcpu
> >> >in this case.
> >>
> >> So unhalt on completion.
> >>
> >I want to avoid touching vcpu state from work if possible. Work code does
> >not contain arch dependent code right now and mp_state is x86 thing
> >
>
> Use a KVM_REQ.
>
Completion happens asynchronously. CPU may not be even halted at that
point. Actually completion does unhalt vcpu. It puts completed work into
vcpu->async_pf.done list and wakes vcpu thread if it sleeps. Next
invocation of kvm_arch_vcpu_runnable() will return true since vcpu->async_pf.done
is not empty and vcpu will be unhalted in usual way by kvm_vcpu_block().
>
> >> >>
> >> >> I guess the apf threads can't touch mp_state, but they can have a
> >> >> KVM_REQ to trigger the check.
> >> >This will require KVM_REQ check on fast path, so what's the difference
> >> >performance wise.
> >>
> >> We already have a KVM_REQ check (if (vcpu->requests)) so it doesn't
> >> cost anything extra.
> >if (vcpu->requests) does not clear req bit, so what will have to be added
> >is: if (kvm_check_request(KVM_REQ_APF_HLT, vcpu)) which is even more
> >expensive then my check (but not so expensive to worry about).
>
> It's only expensive when it happens. Most entries will have the bit clear.
kvm_check_async_pf_completion() (the one that detects if vcpu should be
halted) is called after vcpu->requests processing. It is done in order
to delay completion checking as far as possible in hope to get
completion before next vcpu entry and skip sending apf, so I do it at
the last possible moment before event injection.
> >>
> >> >> >+
> >> >> >+TRACE_EVENT(
> >> >> >+ kvm_async_pf_not_present,
> >> >> >+ TP_PROTO(u64 gva),
> >> >> >+ TP_ARGS(gva),
> >> >>
> >> >> Do you actually have a gva with tdp? With nested virtualization,
> >> >> how do you interpret this gva?
> >> >With tdp it is gpa just like tdp_page_fault gets gpa where shadow page
> >> >version gets gva. Nested virtualization is too complex to interpret.
> >>
> >> It's not good to have a tracepoint that depends on cpu mode (without
> >> recording that mode). I think we have the same issue in
> >> trace_kvm_page_fault though.
> >We have mmu_is_nested(). I'll just disable apf while vcpu is in nested
> >mode for now.
>
> What if we get the apf in non-nested mode and it completes in nested mode?
>
I am not yet sure we have any problem with nested mode at all. I am
looking at it. If we have we can skip prefault if in nested.
> >>
> >> >> >+
> >> >> >+ /* do alloc nowait since if we are going to sleep anyway we
> >> >> >+ may as well sleep faulting in page */
> >> >> /*
> >> >> * multi
> >> >> * line
> >> >> * comment
> >> >> */
> >> >>
> >> >> (but a good one, this is subtle)
> >> >>
> >> >> I missed where you halt the vcpu. Can you point me at the function?
> >> >>
> >> >> Note this is a synthetic halt and must not be visible to live
> >> >> migration, or we risk live migrating a halted state which doesn't
> >> >> really exist.
> >> >>
> >> >> Might be simplest to drain the apf queue on any of the save/restore ioctls.
> >> >>
> >> >So that "info cpu" will interfere with apf? Migration should work
> >> >in regular way. apf state should not be migrated since it has no meaning
> >> >on the destination. I'll make sure synthetic halt state will not
> >> >interfere with migration.
> >>
> >> If you deliver an apf, the guest expects a completion.
> >>
> >There is special completion that tells guest to wake all sleeping tasks
> >on vcpu. It is delivered after migration on the destination.
> >
>
> Yes, I saw.
>
> What if you can't deliver it? is it possible that some other vcpu
How can this happen? If I can't deliverer it I can't deliver
non-broadcast apfs too.
> will start receiving apfs that alias the old ones? Or is the
> broadcast global?
>
Broadcast is not global but tokens are unique per cpu so other vcpu will
not be able to receiving apfs that alias the old ones (if I understand
what you mean correctly).
--
Gleb.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>