GEOM strange error

[Andriy Gapon]

On 13/03/2018 13:52, Eugene Grosbein wrote:
> On 13.03.2018 17:39, Andriy Gapon wrote:
>> On 13/03/2018 11:37, Eugene Grosbein wrote:
>>> Hi!
>>>
>>> Let's create a stripe and GPT over it using test files as backing store:
>>>
>>> truncate -s 1G d0
>>> truncate -s 1G d1
>>> mdconfig -af d0		# gives md0
>>> mdconfig -af d1		# gives md1
>>>
>>> gpart create -s GPT md0
>>> gpart create -s GPT md1
>>> gpart destroy -F md1
>>> gpart destroy -F md0	# no errors still
>>>
>>> gstripe label -s $((128*1024)) st0 md0 md1
>>> gpart create -s GPT /dev/stripe/st0
>>> dmesg -a
>>>
>>> GEOM_STRIPE: Device st0 created (id=4000112224).
>>> GEOM_STRIPE: Disk md0 attached to st0.
>>> GEOM_STRIPE: Disk md1 attached to st0.
>>> GEOM_STRIPE: Device stripe/st0 activated.
>>> GEOM: md0: corrupt or invalid GPT detected.
>>> GEOM: md0: GPT rejected -- may not be recoverable.
>>>
>>> Why does it emit such md0-related error?
>>
>> When GPT is placed on st0 it's opened for writing and, thus, md0 and md1 are
>> open for writing too.  Afterwards, the write access count is cleared from three
>> of them and that triggers re-tasting.  I guess that g_part code tries to taste
>> md0 and md1 and sees the GPT label at the start of md0 and the label is
>> correctly rejected because it's inconsistent with md0's parameters.
> 
> Shouldn't gstripe grab its components for exclusive access?
> So that GEOM does not even try to treat md[01] as targets for re-tasting.

I don't know what it should do, but I see that it doesn't do it in the "idle"
state.  However, when the stripe itself is opened / accessed, then it does add
the exclusive bit to requested access counts.  See g_stripe_access().
I think that what you expected makes more sense for me than what the code
actually does.

-- 
Andriy Gapon