Comment 1 for bug 1876678

Hi Ubuntu / KVM Maintainers,

I have now done additional diagnostics on this bug and it appears to be triggered in nested virtualization case when apic virtualisation is available in Layer 0 HW and then passed forward to Layer 1 VM via Libvirt: <cpu mode='host-model' check='partial'> .

Testing found that in case where Layer 1 FreeBSD host had this feature, see "VID,PostIntr" in "VT-x: PAT,HLT,MTF,PAUSE,EPT,UG,VPID,VID,PostIntr" from CPU Feature below:

<<START LAYER 1 - FreeBSD CPU Report from dmesg.boot>>
...
...
CPU: Intel Core Processor (Broadwell, IBRS) (2600.09-MHz K8-class CPU)
  Origin="GenuineIntel" Id=0x306d2 Family=0x6 Model=0x3d Stepping=2
  Features=0xf83fbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,MMX,FXSR,SSE,SSE2,SS>
  Features2=0xfffa3223<SSE3,PCLMULQDQ,VMX,SSSE3,FMA,CX16,PCID,SSE4.1,SSE4.2,x2APIC,MOVBE,POPCNT,TSCDLT,AESNI,XSAVE,OSXSAVE,AVX,F16C,RDRAND,HV>
  AMD Features=0x2c100800<SYSCALL,NX,Page1GB,RDTSCP,LM>
  AMD Features2=0x121<LAHF,ABM,Prefetch>
  Structured Extended Features=0x1c0fbb<FSGSBASE,TSCADJ,BMI1,HLE,AVX2,SMEP,BMI2,ERMS,INVPCID,RTM,RDSEED,ADX,SMAP>
  Structured Extended Features2=0x4<UMIP>
  Structured Extended Features3=0xac000400<MD_CLEAR,IBPB,STIBP,ARCH_CAP,SSBD>
  XSAVE Features=0x1<XSAVEOPT>
  IA32_ARCH_CAPS=0x8<SKIP_L1DFL_VME>
  AMD Extended Feature Extensions ID EBX=0x1001000
  VT-x: PAT,HLT,MTF,PAUSE,EPT,UG,VPID,VID,PostIntr
Hypervisor: Origin = "KVMKVMKVM"
...
...
<END LAYER 1 - dimes.log>>

In my case with Intel Broadwell chipset this is available, in case of desktop "core i5-8250U" chip- this reports as:

VT-x: PAT,HLT,MTF,PAUSE,EPT,UG,VPID

For this case HW case, nested:
Layer 0 - Ubuntu 20.04, Layer 1 - FreeBSD 12.1 with bhyve, Layer 2 - FreeBSD 12.1
Works.

Workaround is to disable APIC virtual interrupt delivery:

1. Add entry into Layer 1 - FreeBSD Guest / Host: /boot/loader.conf:
hw.vmm.vmx.use_apic_vid=0

2. Reboot

3. Check via sysctl that virtual_interupt_delivery is disabled:
# sysctl hw.vmm.vmx.cap.virtual_interrupt_delivery
hw.vmm.vmx.cap.virtual_interrupt_delivery: 0 <- should be zero

Questions is:

While FreeBSD triggers this bug, is this a KVM issue or a FreeBSD bhyve one ?

In doing some searching on Web I see that there is already work being done with KVM 5.6 around APIC virtualisation and its handling. So not sure if this a potentially know problem: https://events19.linuxfoundation.org/wp-content/uploads/2017/12/Improving-KVM-x86-Nested-Virtualization-Liran-Alon-Oracle.pdf

APIC Virtualisation support was introduced back in FreeBSD 11.0 way back in Sept 2016:

https://www.freebsd.org/releases/11.0R/relnotes.html#hardware-virtualization

Thanks to Peter Graham on FreeBSD virtualization bug tracker for helping to find source of problem.

Should this BUG go to KVM / QEMU upstream ?

Cheers,

John Hartley.