Re: Re: Oracle RAC in libvirt+KVM environment

[Masaki Kimura <masaki.kimura.kz@hitachi.com>]

>> I found that scsicmd can't pass all the scsi3_test but the result of
>> sg_inq is the same as it in the host.
>>
>> I am absolutely confused about this situation. Am I missed some
>> information about it?

I guess this is caused by the lack of capability.
Please check if enough capability was added to kvm process by the following 
steps. 

  1. Check the pid of kvm process.
     # ps -C qemu-system-x86_64 -o pid=
      5177
  2. Check the capability for the process.
     # getpcaps 5177
     Capabilities for `5177': = cap_sys_rawio+i

In my fedora19 environment, as seen in above, only cap_sys_rawio+i was added
with rawio='yes'. Even though, cap_sys_rawio+ep is required to pass-through 
SCSI Reservation from the guest. 

Note that I succeeded to pass-through SCSI Reservation with the following steps
in my environment, not a Windows guest though.

  1. Stop the guest.
  2. Add CAP_SYS_RAWIO(effective, permitted) to qemu-kvm.
     # setcap cap_sys_rawio=ep /usr/bin/qemu-system-x86_64
  3. Start the guest.

However, I don't think this is the right way to workaround it, because it 
gives cap_sys_rawio+ep to all the kvm processes executed from this binary.

I believe following patches, which are not merged yet, are trying to solve 
this problem in a different approach.
     - [PATCH v3 part2] Add per-device sysfs knob to enable unrestricted, 
      unprivileged SG_IO
      https://lkml.org/lkml/2013/5/23/294
    - [RFC PATCH 0/4] SG_IO filtering via sysfs and minimal whitelist
      https://lkml.org/lkml/2013/5/27/230


Any comments on this?


Masaki