[libvirt] [PATCH v2 3/3] disk: Force clearing of the first 2048 bytes of device prior to mklabel

[John Ferlan]

On 04/07/2017 10:31 AM, Daniel P. Berrange wrote:
> On Fri, Apr 07, 2017 at 04:27:40PM +0200, Peter Krempa wrote:
>> On Fri, Apr 07, 2017 at 09:52:09 -0400, John Ferlan wrote:
>>> https://bugzilla.redhat.com/show_bug.cgi?id=1439132
>>>
>>> During 'matrix' testing of all possible combinations I found that if
>>> device is formated with "gpt" first, then an attempt is made to format
>>> using "mac", a startup will fail.  By adding a clearing of the first
>>> 2048 bytes of the device (similar to the logical pool code), the issue
>>> is resolved.
>>
>> I'm not a fan of this and neither thing this is robust enough:
>>
>> The problem here is that apparently the "mac" table fits into the first
>> block on the disk. Since the GPT disklabel is stored at LBA address 1
>> it is not overwritten at all. Thus it's apparent that the detection tool
>> then prefers GPT over a older disklabel.
>>
>> The GPT disklabel has also a secondary copy at the last LBA of the disk.
>>
>> As for robustness: SECTOR_SIZE is defined as 512 in our code, thus this
>> code nukes 2k of data. For devices with 4k sectors which are already
>> available this won't help at all, since LBA 0 will still contain the
>> protective MBR, and the 3,5k left of LBA 0 will be unused. GPT disklabel
>> still will start at LBA 1.
> 
> Personally I'd forget about sectors and just nuke 1 MB of data at start
> and end of the device. The time for 2k vs 1 MB is negligible.
> 
> 

OK - so I've pushed patches 4 & 5 and will rework the first 3 ... series
coming soon the a mailing list near you.

John