Re: [Qemu-devel] [BUG] Migrate failes between boards with different PMC counts

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Zhuangyanying (ann.zhuangyanying@huawei.com) wrote:
> Hi all,
> 
> Recently, I found migration failed when enable vPMU.
> 
> migrate vPMU state was introduced in linux-3.10 + qemu-1.7.
> 
> As long as enable vPMU, qemu will save / load the
> vmstate_msr_architectural_pmu(msr_global_ctrl) register during the migration.
> But global_ctrl generated based on cpuid(0xA), the number of general-purpose performance
> monitoring counters(PMC) can vary according to Intel SDN. The number of PMC presented
> to vm, does not support configuration currently, it depend on host cpuid, and enable all pmc
> defaultly at KVM. It cause migration to fail between boards with different PMC counts.
> 
> The return value of cpuid (0xA) is different dur to cpu, according to Intel SDN，18-10 Vol. 3B:
> 
> Note: The number of general-purpose performance monitoring counters (i.e. N in Figure 18-9)
> can vary across processor generations within a processor family, across processor families, or
> could be different depending on the configuration chosen at boot time in the BIOS regarding
> Intel Hyper Threading Technology, (e.g. N=2 for 45 nm Intel Atom processors; N =4 for processors
> based on the Nehalem microarchitecture; for processors based on the Sandy Bridge
> microarchitecture, N = 4 if Intel Hyper Threading Technology is active and N=8 if not active).
> 
> Also I found, N=8 if HT is not active based on the broadwell，,
> such as CPU E7-8890 v4 @ 2.20GHz   
> 
> # ./x86_64-softmmu/qemu-system-x86_64 --enable-kvm -smp 4 -m 4096 -hda
> /data/zyy/test_qemu.img.sles12sp1 -vnc :99 -cpu kvm64,pmu=true -incoming tcp::8888
> Completed 100 %
> qemu-system-x86_64: error: failed to set MSR 0x38f to 0x7000000ff
> qemu-system-x86_64: /data/zyy/git/test/qemu/target/i386/kvm.c:1833: kvm_put_msrs: 
> Assertion `ret == cpu->kvm_msr_buf->nmsrs' failed.
> Aborted
> 
> So make number of pmc configurable to vm ? Any better idea ?

Coincidentally we hit a similar problem a few days ago with -cpu host  - it took me
quite a while to spot the difference between the machines was the source
had hyperthreading disabled.

An option to set the number of counters makes sense to me; but I wonder
how many other options we need as well.  Also, I'm not sure there's any
easy way for libvirt etc to figure out how many counters a host supports - it's
not in /proc/cpuinfo.

Dave

> 
> Regards,
> -Zhuang Yanying
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK