geom_raid tasting providers that can't be raw disks

[Alexander Motin]

Gavin Atkinson wrote:
> On Thu, 23 Jun 2011, Alexander Motin wrote:
>> On 23.06.2011 20:53, Gavin Atkinson wrote:
>>> While debugging a problem that looks like it was unrelated to geom_raid,
>>> I realised that it tastes all providers, including each partition and
>>> slice on raid devices it itself created.
>>>
>>> Given that geom_raid is purely a replacement for ataraid, the only place
>>> that the metadata can be valid is on the raw disk itself.
>> In general case nothing prevents from using graid on partitions (instead of
>> gmirror/gstripe/...) if BIOS support is not needed. I think this check at
>> least should be moved to specific metadata modules in case if we later add
>> support for abstract gxxx metadata formats.
> 
> OK, thanks.  I had believed that this was purely for use on hardware.  If 
> the intention is that this module can be used on arbitrary providers, then 
> I see no reason to change things.

It may depend on metadata format. For example, Intel format depends on
valid provider serial number. It can't be used on providers without
serial number -- it won't be able to find place in array for the
components. Not sure if any non-hardware provider has serial number, so
this limitation can be effectively the same.

>>> Also, should geom_raid be in GENERIC?  ataraid was, and it's one less
>>> "gotcha" for upgrades.  Given the lack of ar0 ->  raid/r0 aliases, the
>>> upgrade is painful enough for users already, putting it in GENERIC may
>>> at least help slightly...
>> Aggressive tasting for each metadata format was actually the reason why I
>> haven't added it. If we load all GEOM modules, then some floppy tasting will
>> take ages.
> 
> OK, that makes sense.  
> 
> I had a bit of a look this evening as to the best way to try to get code 
> in place to provide /dev/arX compatibility nodes whenever a geom_raid 
> array is discovered.  Is there any reason that I shouldn't take exactly 
> the same approach as you have done with the adX -> adaX compatibility 
> shims?

May be not. Just haven't tried. Should be only few lines.

-- 
Alexander Motin