Stefan Smietanowski wrote:
> Hi Steve.
>
>>> Yup. That's what's happening. It first does one run with --just-copy
>>> where it writes the files using the filesystem then reads the same
>>> files using the blockdevice and it's own filesystem code basically.
>>>
>>> // Stefan
>>>
>>
>> And I presume the files are missing?
>>
>> The bizzare part of this is that if you read the via the block
>> device interface, you are looking at the same in memory pages
>> which xfs uses for the metadata cache. So even if the data has
>> not hit disk yet, things such as names in directories should
>> be visible in the metadata cache. Inodes may be more tricky,
>> since flushing of inodes into the metadata cache is delayed.
>>
>> Try this for an experiment, before the run, set
>>
>> /proc/sys/fs/xfs/sync_interval
>>
>> down to some small number like 1000, pause for a couple
>> of seconds after calling sync, and then see if grub
>> can see the files via the block device.
>>
>> This tunable controls how long xfs delays writing out
>> inodes into the metadata cache from their internal format.
>
>
> Alright. I'll try that on my DVD in a few hours and get back to you.
>
> Just to verify to be accurate : That /proc entry exists regardless of
> if xfs is statically compiled or as a module, correct? (ATRPMS makes
> it a module).
>
> // Stefan
>
It will exist if you have an xfs filesystem and /proc mounted. Loading
an xfs module will create it. Looks like both 2.4 and 2.6 use this now.
Steve
|