Re: [PATCH v2 3/3] btrfs-progs: device: add remove missing-all

[Anand Jain <anand.jain@oracle.com>]

On 10/13/2017 01:27 PM, Duncan wrote:
> Misono, Tomohiro posted on Wed, 11 Oct 2017 11:18:50 +0900 as excerpted:
> 
>> Add 'btrfs remove missing-all' to remove all the missing devices at once
>> for improving usability.
>>
>> Example:
>>   sudo mkfs.btrfs -f -d raid1 /dev/sdb1 /dev/sdb2 /dev/sdb3 /dev/sdb4
>>   sudo wipefs -a /dev/sdb1 /dev/sdb3
>>   sudo mount -o degraded /dev/sdb2 /mnt <-- 


   I agree with Duncan here. This step itself will fail even with RO
   option. Do you have any patch that is not in the ML which will
   make this step a success in the first place ?

Thanks, Anand


>>   sudo btrfs filesystem show /mnt
>>   sudo btrfs device remove missing-all /mnt
>>   sudo btrfs filesystem show /mnt
> 
> 
> There's a reason remove missing-all hasn't yet been implemented.
> 
> Note that the above would be very unlikely to work once a filesystem has
> been used in any significant way, because raid1 and raid10 are explicitly
> chunk pairs, *NOT* duplicated N times across N devices.  So with two
> devices missing, chances are that both copies of some chunks will be
> missing as well, so the filesystem would no longer be mountable degraded-
> writable, only degraded-readonly, in which case device remove won't work
> at all because the filesystem is readonly.
> 
> In fact, until the recent per-chunk check patches went in, it was
> impossible to mount-writable a raid1 missing two devices at all, because
> the safeguards simply assumed some chunks would be entirely missing.
> 
> The only case in which more than a single device missing is likely to be
> mountable degraded-writable (so device remove will work at all) is raid6,
> tho with recent patches there's narrow cases in which it /might/ be
> doable with raid1 as well.
> 
> Now you may still wish to implement remove missing-all for raid6 mode and
> for the unusual corner-case raid1/raid10 in which it might work, but the
> documentation should be pretty clear that save for raid6 it can't be
> expected to work in most cases.
> 
> Given that, I think remove missing-all hasn't been implemented as it
> simply hasn't been considered to be worth the bother for the narrow use-
> cases in which it will actually work.
>