Skip to content

chapter-16.md

Latest commit

 

History

History
73 lines (42 loc) · 2.07 KB

chapter-16.md

File metadata and controls

73 lines (42 loc) · 2.07 KB

Chapter.16 Hypervisors

Background

  • Hypervisor Patterns
  • Xen Style: a native (bare-metal) hypervisor running directly on processors. Create domains for VMs.
  • KVM style: hypervisor run/administered by Host OS kernel. H/w access via Kernel Modules.

  • Xen style, runs I/O proxy (like QEMU) in domain0. KVM style on host OS.

  • AMD-V & Intel VT-x extensions bring processor virtualization support.

  • Paravirtualization has guest OS aware of virtual machine, making hypercalls to hypervisor for efficiency. Xen batches these calls into multicall.

Now, Xen VMs boot in H/w VM (HVM), then use PV.. hence PVHVM.

  • @Guests: BPF to check virtualized h/w perf; hypercall latency; stolen CPU time; hypervisor interrupt callbacks.

  • @Hosts: Instrument VM exits; I/O proxy workload & latency; prior tools for perf.

Traditional Tools

  • Guests sometimes have tracepoints for Xen Hypercalls.

  • Hosts have top/trace tools for hypervisor provided, like xl top, xentrace.

Guest BPF Tools

  • Xen Hypercalls using funccount, trace, argdist, stackcount; get PV status via dmesg | grep Hypervisor. Like funccount 't:xen:*'.
argdist -C 't:xen:xen_mc_flush():int:args->mcidx'  ## counting Hypercalls

stackcount 't:xen:xen_mc_issue'  ## hypercall stacks

funclatency xen_mc_flush  ## hypercall latency

  • xenhyper, bpftrace tool counts hyperalls via xen:xen_mc_entry tp.

  • When Xen calls guest, /proc/interrupts get grep-able HYP Xen callback entries. Trace via bpftrace -e 'kprobe:xen_evtchn_do_upcall { @[comm] = count(); }'.

  • cpustolen, bpftrace tool list stolen CPU time dist. Cab also include time in VMM on behalf of guest. Similar code

#!bpftrace

BEGIN {
    printf("Tracing stolen CPU time. Ctrl-C to end.\n");
}

kretprobe:xen_steal_clock, kretprobe:kvm_steal_clock {
    if (@last[cpu] > 0) {
        @stolen_us = hist((retval - @last[cpu]) / 1000);
    }
    @last[cpu] = retval;
}

END {
    clear(@last);
}

Host BPF Tools

  • kvmexits, bpftrace tool to show guest exit time dist. Works by tracing kvm:kvm_exit & kvm:kvm_entry.