Sean Christopherson [off-list ref] writes:

On Tue, May 17, 2022, Vitaly Kuznetsov wrote:

quoted

Sean Christopherson [off-list ref] writes:

quoted

On Thu, Apr 14, 2022, Vitaly Kuznetsov wrote:

quoted

To make kvm_hv_flush_tlb() ready to handle L2 TLB flush requests, KVM needs
to allow for all 64 sparse vCPU banks regardless of KVM_MAX_VCPUs as L1
may use vCPU overcommit for L2. To avoid growing on-stack allocation, make
'sparse_banks' part of per-vCPU 'struct kvm_vcpu_hv' which is allocated
dynamically.

Note: sparse_set_to_vcpu_mask() keeps using on-stack allocation as it
won't be used to handle L2 TLB flush requests.

I think it's worth using stronger language; handling TLB flushes for L2 _can't_
use sparse_set_to_vcpu_mask() because KVM has no idea how to translate an L2
vCPU index to an L1 vCPU.  I found the above mildly confusing because it didn't
call out "vp_bitmap" and so I assumed the note referred to yet another sparse_banks
"allocation".  And while vp_bitmap is related to sparse_banks, it tracks something
entirely different.

Something like?

Note: sparse_set_to_vcpu_mask() can never be used to handle L2 requests as
KVM can't translate L2 vCPU indices to L1 vCPUs, i.e. its vp_bitmap array
is still bounded by the number of L1 vCPUs and so can remain an on-stack
allocation.

My brain is probably tainted by looking at all this for some time so I
really appreciate such improvements, thanks :)

I wouldn't, however, say "never" ('never say never' :-)): KVM could've
kept 2-level reverse mapping up-to-date:

KVM -> L2 VM list -> L2 vCPU ids -> L1 vCPUs which run them

making it possible for KVM to quickly translate between L2 VP IDs and L1
vCPUs. I don't do this in the series and just record L2 VM_ID/VP_ID for
each L1 vCPU so I have to go over them all for each request. The
optimization is, however, possible and we may get to it if really big
Windows VMs become a reality.

Out of curiosity, is L1 "required" to provides the L2 => L1 translation/map?

To make this "Direct Virtual Flush" feature work? Yes, it is:

...
"
Before enabling it, the L1 hypervisor must configure the following
additional fields of the enlightened VMCS:
- VpId: ID of the virtual processor that the enlightened VMCS controls.
- VmId: ID of the virtual machine that the enlightened VMCS belongs to.
- PartitionAssistPage: Guest physical address of the partition assist
page.
"

-- 
Vitaly