[libvirt] [PATCH 0/2] Resolve issues seen with schedinfo

[John Ferlan]

ping?  Although perhaps I should have noted in the title that it seems
there's a regression as a result of cgroups changes to the
virGetSchedulerType() and virDomainGetSchedulerParameters[Flags]() APIs

John

On 05/30/2013 08:24 AM, John Ferlan wrote:
> These patches resolve an issue seen using 'virsh schedinfo <domain>' on a
> non running domain that have been present since 1.0.4 as a result of the
> cgroup infrastructure changes:
> 
> https://www.redhat.com/archives/libvir-list/2013-April/msg00783.html
> 
> The exact commit id that caused the issue is listed in each of the commit
> messages.  I used git bisect to determine, although it was tricky because
> the TPM changes were made around the same time and required commit '8b934a5c'
> to be applied in order to actually see domains on my host.
> 
> Prior to the changes the "CFS Bandwidth" data provided by qemuGetCpuBWStatus()
> and lxcGetCpuBWStatus() would be obtained as a result of the driver cgroup
> being mounted. Now it relies on the domain cgroup being mounted which only
> occurs once the domain is active.
> 
> This issue also affects the libvirt-cim code in how it defines QEMU domains.
> Fortunately it only looks for the "cpu_shares" value.
> 
> A "downside" to the end result of these changes is that for a non-running
> domain it becomes impossible to obtain the vcpu_period, vcpu_quota,
> emulator_period, and emulator_quota values. All that can be obtained is
> the 'cpu_shares" value. As an option, I did consider adding the following to
> either [qemu|lxc]DomainSchedulerGetType() and DomainSchedulerParameter[Flags]()
> or just [qemu|lxc]GetCpuBWStatus() APIs in order to get the "host" values if
> they existed (since they are eventually copied into the domain cgroup):
> 
>     if (cgroup == NULL) {
>         rc = virCgroupNewSelf(&hostgrp);
>         if (rc < 0) {   
>             virReportSystemError(-rc, "%s",
>                                  _("Unable to get schedinfo host cgroup"));
>             goto cleanup;                           
>         }                                                   
>         cgroup = hostgrp;       
>     }                               
> 
> However, I wasn't sure going that route was desired and figured that I'd
> use the code review opportunity to get the answer. Furthermore, it wasn't
> clear that the vcpu_* and emulator_* values made sense for a non running
> domain. Also, the virReportSystemError may need to change to a VIR_INFO
> since I believe it would be triggered if cgroups weren't mounted on the system.
> 
> Another option would be to just add the above code to the GetType() APIs and
> then ignore the 'cpu_bw_status' for the VIR_DOMAIN_AFFECT_CONFIG path in
> the ParametersFlags() APIs thus returning all 5 datapoints set to whatever
> the persistentDef had defined.
> 
> Since this has been present since 1.0.4 and no one has complained so far, I
> don't think it's critical for 1.0.6. I suspect the change would need to into
> the maint trees though. It might be nice to get DanB's opinion on this too.
> 
> John Ferlan (2):
>   qemu: Resolve issue with virsh schedinfo for non running domain
>   lxc: Resolve issue with virsh schedinfo for non running domain
> 
>  src/lxc/lxc_driver.c   |  9 ++++++++-
>  src/qemu/qemu_driver.c | 11 ++++++++++-
>  2 files changed, 18 insertions(+), 2 deletions(-)
>