[et-mgmt-tools] Virtinst and blktap

[Michael DeHaan]

Cole Robinson wrote:
> Michael DeHaan wrote:
>   
>> Cole Robinson wrote:
>>     
>>> Michael DeHaan wrote:
>>>   
>>>       
>>>> Chris Lalancette wrote:
>>>>     
>>>>         
>>>>> Michael DeHaan wrote:
>>>>>   
>>>>>       
>>>>>           
>>>>>> I had a user question about how they can specify blktap for Xen in 
>>>>>> Cobbler instead of the file driver (for performance reasons) ... well, 
>>>>>> it turns out they can't (yet).
>>>>>>
>>>>>> Under what conditions can we use blktap, and is that advisable?   (I had 
>>>>>> thought this might work like virtio, in which case, passing in the os 
>>>>>> version uses a table, but it seems this may only actually depend on the 
>>>>>> host and has no guest limitations?)
>>>>>>     
>>>>>>         
>>>>>>             
>>>>> That's exactly right.  It's a property of how the host OS accesses the disk on
>>>>> behalf of the domain, not of how the disk is presented to the domain.  In fact,
>>>>> for Xen PV domains, you should *only* ever use blktap (i.e. tap:aio:); direct
>>>>> file: access can corrupt your guest domains on host crash.  The situation with
>>>>> FV domains is more complicated, but there, you can use file:, since it properly
>>>>> does flushing.
>>>>>
>>>>>   
>>>>>       
>>>>>           
>>>>     
>>>>         
>>>>>> It looks like from the virtinst code I could just call 
>>>>>> utils.is_blktap_capable() from my libvirt-using application and if so, 
>>>>>> then pass the driver=DRIVER_TAP options to virstinst.
>>>>>>
>>>>>> Good idea?  Bad idea?
>>>>>>     
>>>>>>         
>>>>>>             
>>>>> The only reasonable default for Xen PV domains, in my opinion.
>>>>>
>>>>>   
>>>>>       
>>>>>           
>>>> Thanks.
>>>>
>>>> FWIW the default behavior of python-virtinst seems to be to /not/ use 
>>>> blktap, and should probably be fixed then if there's a potential 
>>>> corruption problem.
>>>>
>>>> I haven't tested command line virt-install to see what it does.
>>>>
>>>>     
>>>>         
>>> That default is actually set in virt-install and virt-manager, but
>>> probably should have been moved into the virtinst library a long time ago.
>>>
>>> - COle
>>>   
>>>       
>> Yeah, I can't keep up very well when stuff doesn't happen there :)
>>
>> Should I make the fix in koan also or do you want to move it?
>>
>> --Michael
>>
>>     
>
> I'll take a look at moving it into the library, but for back compat it
> wouldn't hurt to set the default in koan. The virt-install code is:
>
> http://hg.et.redhat.com/cgi-bin/hg-virt.cgi/applications/virtinst--devel/file/5d6dc8af58b5/virt-install#l209
>
> - Cole
>   

Excellent, will do.