Re: [PATCH v2 2/5] KVM: x86: hyperv: introduce vp_index_to_vcpu_idx mapping

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Roman Kagan <rkagan@virtuozzo.com> writes:

> On Fri, Jun 29, 2018 at 12:26:23PM +0200, Vitaly Kuznetsov wrote:
>> Roman Kagan <rkagan@virtuozzo.com> writes:
>> 
>> > On Thu, Jun 28, 2018 at 03:53:10PM +0200, Vitaly Kuznetsov wrote:
>> >> While it is easy to get VP index from vCPU index the reverse task is hard.
>> >> Basically, to solve it we have to walk all vCPUs checking if their VP index
>> >> matches. For hypercalls like HvFlushVirtualAddress{List,Space}* and the
>> >> upcoming HvSendSyntheticClusterIpi* where a single CPU may be specified in
>> >> the whole set this is obviously sub-optimal.
>> >> 
>> >> As VP index can be set to anything <= U32_MAX by userspace using plain
>> >> [0..MAX_VP_INDEX] array is not a viable option. Use condensed sorted
>> >> array with logarithmic search complexity instead. Use RCU to make read
>> >> access as fast as possible and maintain atomicity of updates.
>> >
>> > Quoting TLFS 5.0C section 7.8.1:
>> >
>> >> Virtual processors are identified by using an index (VP index). The
>> >> maximum number of virtual processors per partition supported by the
>> >> current implementation of the hypervisor can be obtained through CPUID
>> >> leaf 0x40000005. A virtual processor index must be less than the
>> >> maximum number of virtual processors per partition.
>> >
>> > so this is a dense index, and VP_INDEX >= KVM_MAX_VCPUS is invalid.  I
>> > think we're better off enforcing this in kvm_hv_set_msr and keep the
>> > translation simple.  If the algorithm in get_vcpu_by_vpidx is not good
>> > enough (and yes it can be made to return NULL early on vpidx >=
>> > KVM_MAX_VCPUS instead of taking the slow path) then a simple index array
>> > of KVM_MAX_VCPUS entries should certainly do.
>> 
>> Sure, we can use pre-allocated [0..KVM_MAX_VCPUS] array instead and put
>> limits on what userspace can assign VP_INDEX to. Howver, while thinking
>> about it I decided to go with the more complex condensed array approach
>> because the tendency is for KVM_MAX_VCPUS to grow and we will be
>> pre-allocating more and more memory for no particular reason (so I think
>> even 'struct kvm_vcpu *vcpus[KVM_MAX_VCPUS]' in 'struct kvm' will need
>> to be converted to something else eventually). 
>
> We're talking of kilobytes here.  I guess this is going to be the least
> of the scalability problems.

Yes, kilobytes but per-VM.

>
>> Anyway, I'm flexible and if you think we should go this way now I'll do
>> this in v3. We can re-think this when we later decide to raise
>> KVM_MAX_VCPUS significantly.
>
> Although there's no strict requirement for that I think every sensible
> userspace will allocate VP_INDEX linearly resulting in it being equal to
> KVM's vcpu index.  So we've yet to see a case where get_vcpu_by_vpidx
> doesn't take the fast path.  If it ever starts appearing in the profiles
> we may consider optimiziing it but ATM I don't even think introducing
> the translation array is justified.

It was Radim who suggested it in the first place :-)

The problem we're trying to solve here is: with PV TLB flush and IPI we
need to walk through the supplied list of VP_INDEXes and get VCPU
ids. Usually they match. But in case they don't we'll fall back to full
scan for every VP_INDEX in the supplied list. Now let's say we have 128
CPUs. We'll need to perform up to 128 * 128 extra comparisons on every
hypercall. Not good. So instead of using get_vcpu_by_vpidx() I opted for
walking the whole VCPU list and checking if VPU's VP_INDEX is in the
supplied set. This way we end up with 128 comparisons in the example
above (worst case scenarion). However, we lose in simple scenarios like
only 1 VP_INDEX was specified in the set: we'll still need to walk the
whole list. So having the translation array (one way or another) is IMO
justified.

-- 
  Vitaly