Re: [Qemu-devel] Re: KVM call agenda for Jan 11

[Anthony Liguori <anthony@codemonkey.ws>]

On 01/11/2011 07:41 AM, Juan Quintela wrote:
> Kevin Wolf<kwolf@redhat.com>  wrote:
>    
>> Am 10.01.2011 14:32, schrieb Juan Quintela:
>>      
>>> Juan Quintela<quintela@redhat.com>  wrote:
>>>        
>>>> Juan Quintela<quintela@redhat.com>  wrote:
>>>>
>>>> Now sent it to the right kvm list.  Sorry for the second sent.
>>>>
>>>>          
>>>>> Please send any agenda items you are interested in covering.
>>>>>
>>>>> - KVM Forum 2011 (Jes).
>>>>>
>>>>> thanks, Juan.
>>>>>            
>>> - migration and block devices: a mess.
>>>    * patches I sent last week: only work for root (for some definition of
>>>      work)
>>>    * qemu is used as non-root user.
>>>    * forcing to have cache=none solves the issue
>>>        
>> I need to have a look at the specific problem, but it's hard to imagine
>> that cache=none fixes anything reliably.
>>      
> It uses O_DIRECT, that means that we don't have buffering problems.
> I state the problem again:
>
> machine A read 1st block of device.
> <and stays without doing anything else>
> machine B reads writes lots of places including 1st block
>
> now guest from machine A migrates to machine B
> machine A re-reads the 1st block, and lo and behold, it reads the old
> contents, not the new ones.
>
> Solutions:
> - invalidate all buffers for that block device on machine A after
>    migration.
>     * with NFS, just close + reopen the file (and pray that nobody else
>     has it also opened)
>     * with block devices: use BLKFLBLK ioctl, and pray that nobody else is
>       using the device, that device is not a ramdisk, and some more
>       things.  To add injury to insult, you need to be root to be able
>       to issue that ioctl (technically have CAP_SYS_ADMIN).
>    

Why isn't fsync() enough for a block device?

Regards,

Anthony Liguori

> O_DIRECT fixes this problem altogether, because there is no buffering,
> and if there are not buffers, they can't be invalid O:-)
>
> Notice the "pray" part in the other solutions, we are basically trying
> to do a "poor man" DLM, and that is not trivial to do. (althougth our
> problem is not the general one, the principles are the same).
>
> Later, Juan.
>
>
>