Re: [PATCH] rootfs.py: Remove rpm database from staging area

[Mark Hatle <mark.hatle@windriver.com>]

On 3/30/15 3:53 AM, Richard Purdie wrote:
> On Mon, 2015-03-30 at 11:30 +0300, Ed Bartosh wrote:
>> Rpm database in staging area is used only by createrepo.
>> createrepo fails with the error
>> "rpmdb: BDB0060 PANIC: fatal region error detected"
>> if rpm database is broken from the previous run of createrepo.
>>
>> Removing the databae before running createrepo can hopefully
>> prevent this failure to happen.
>>
>> [YOCTO #6571]
>>
>> Signed-off-by: Ed Bartosh <ed.bartosh@linux.intel.com>
>> ---
>>  meta/lib/oe/rootfs.py | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/meta/lib/oe/rootfs.py b/meta/lib/oe/rootfs.py
>> index 4e4e6eb..9f7dc65 100644
>> --- a/meta/lib/oe/rootfs.py
>> +++ b/meta/lib/oe/rootfs.py
>> @@ -306,6 +306,9 @@ class RpmRootfs(Rootfs):
>>              bb.utils.remove(self.image_rootfs, True)
>>          else:
>>              self.pm.recovery_packaging_data()
>> +        dbpath = os.path.join(self.d.getVar('STAGING_DIR_NATIVE', True),
>> +                              'var/lib/rpm/*')
>> +        bb.utils.remove(dbpath, recurse=True)
>>          bb.utils.remove(self.d.getVar('MULTILIB_TEMP_ROOTFS', True), True)
>>  
>>          self.pm.crea
> 
> This patch helps me see the problem a lot. I'd never realised there was
> an rpm database in the native sysroot that was getting corrupted, I'd
> always assumed it was the target rootfs one.
> 
> I'm a little worried about what happens if you have multiple images
> generating at the same time as this change as above may delete something
> being worked on by another process.
> 
> I'm wondering why is it getting into the sysroot at all? Rather than
> delete it here, could we either not generate it at all, or place it
> somewhere in WORKDIR (so that other tasks can't see it or race against
> it)?

My guess is that either createrepo or other rpm commands are doing a lazy init
on the DB.  As you mentioned deleting the DB is a little worrying because if
it's open in another process you -are- going to cause other failures.

It would be nice if we could find the creation points and redirect them to the
current workdir.  (Pseudo and it's logging might help us be able to do that.)
The other option could be to create an empty DB, and simply set it and the
directory containing it to read-only, no write.... (but I'm not sure that the
filesystem is actually going to enforce that for the current user.)

--Mark

> Cheers,
> 
> Richard
>