Re: [regression] Bug 217074 - upgrading to kernel 6.1.12 from 5.15.x can no longer assemble software raid0

[Mariusz Tkaczyk]

On Mon, 06 Mar 2023 08:21:07 +1100
"NeilBrown" <neilb@xxxxxxx> wrote:

> On Sat, 04 Mar 2023, Song Liu wrote:
> > + Jes.
> > 
> > It appeared to me that we can assemble the array if we have any of the
> > following:
> > 1. Enable CONFIG_BLOCK_LEGACY_AUTOLOAD;
> > 2. Have a valid /etc/mdadm.conf;
> > 3. Update mdadm to handle this case. (I tried some ugly hacks, which worked
> > but weren't clean).
> > 
> > Since we eventually would like to get rid of CONFIG_BLOCK_LEGACY_AUTOLOAD, I
> > think we need mdadm to handle this properly. But the logistics might
> > be complicated, as
> > mdadm are shipped separately.
> > 
> > Jes, what do you think about this? AFAICT, we need to update the logic in
> > mdopen.c:create_mddev().  
> 
> mdadm already handles this, but only if 
>    CREATE names=yes
> is present in /etc/mdadm.conf

Hi,

"CREATE names=yes" enforces creation of /dev/md_name arrays instead of
/dev/mdXXX. It is a large change for users, too aggressive IMO. It will destroy
many setups.

To resolve it, we need is to use create_named_array() but respect old naming
convention. We already have find_free_devnm(), and we are able to keep
consistency because we can create 'mdXXX':

/sys/module/md_mod/parameters # echo md125 > new_array

/sys/module/md_mod/parameters # ll /sys/block/md125
lrwxrwxrwx 1 root root 0 Mar  7 10:54 /sys/block/md125 ->
../devices/virtual/block/md125

That will require adjustments in mdadm, but I think that we can keep
names the same way. I created the test for verification of base creation flows,
we can use it to avoid regression:
https://git.kernel.org/pub/scm/utils/mdadm/mdadm.git/tree/tests/00createnames

Thoughts?

BTW. I wanted to get rid of this legacy "create_on_open" from mdadm anyway but
never had time to. If you agree, I can proceed with fixing it.

Thanks,
Mariusz