On Tue, 13 Feb 2018 18:18:45 +0100 Viktor Mihajlovski <mihaj...@linux.vnet.ibm.com> wrote:
> This series consolidates patches around a performance issue > caused by the usage of QMP query-cpus. > > A performance issue was found in an OpenStack environment, where > ceilometer was collecting domain statistics with libvirt. The domain > statistics reported by libvirt include the vCPU halted state, which > in turn is retrieved with QMP query-cpus. > > This causes two issues: > 1. Performance: on most architectures query-cpus needs to issue a KVM ioctl > to find out whether a vCPU was halted. This is not the case for s390 > but query-cpus is always causing the vCPU to exit the VM. > > 2. Semantics: on x86 and other architectures, halted is a highly transient > state, which is likely to have already changed shortly after the state > information has been retrieved. This is not the case for s390, where > halted is an indication that the vCPU is stopped, meaning its not > available to the guest operating system until it has been restarted. > > The following patches help to alleviate the issues: > > Patch 1/3: > Adds architecture specific data to the QMP CpuInfo type, exposing > the existing s390 cpu-state in QMP. The cpu-state is a representation > more adequate than the ambiguous 'halted' condition. > > Patch 2/3: > Adds a new QMP function query-cpus-fast, which will only retrieve > vCPU information that can be obtained without interrupting the > vCPUs of a running VM. It introduces a new return type CpuInfoFast > with the subset of fields meeting this condition. Specifically, the > halted state is not part of CpuInfoFast. QMP clients like libvirt > are encouraged to switch to the new API for vCPU information. > > Patch 3/3: > Adds the s390-specific cpu state to CpuInfoFast, allowing management > apps to find out whether a vCPU is in the stopped state. This extension > leads to a partial duplication of field definitions from CpuInfo > to CpuInfoFast. This should be tolerable if CpuInfo is deprecated and > eventually removed. How shall we proceed with this series? Patch 3 depends upon patch 1, so I think it makes sense to merge this in one go. I can give my R-b on patch 1 and Someone(tm) can merge this, or I can take the whole series through the s390 tree (with some further reviews/acks on patches 2/3).