Re: [RFC ONLY 3/8] btrfs: read raid-stripe-tree from disk

[Qu Wenruo <wqu@suse.com>]

On 2022/5/19 16:39, Johannes Thumshirn wrote:
> On 19/05/2022 10:36, Qu Wenruo wrote:
>>
>>
>> On 2022/5/18 19:29, Johannes Thumshirn wrote:
>>> On 17/05/2022 10:28, Qu Wenruo wrote:
>>>>>
>>>>> Metadata on the stripe tree really is the main blocker right now.
>>>>
>>>> That's no doubt.
>>>
>>> What could be done and I think this is the only way forward, is to have
>>> the stripe tree in the system block group
>>
>> This behavior itself has its problems, unfortunately.
>>
>> Currently the system chunks are pretty small, in fact system chunks has
>> the minimal stripe size for current code base.
>> (Data: 1G, Meta: 1G/256M, sys: 32M)
>>
>> This means, if we put stripe tree (which can be as large as extent tree
>> afaik), we need way much larger system chunks.
> 
> I know, but IIRC (need to look this up again) Josef increased the max size
> of sys chunks to 2G
> 
>>
>> And this can further increase the possibility on ENOSPC due to
>> unbalanced data/metadata/sys usage.
>>
>> Although this is really the last problem we need to bother.
>>
>>> and force system to be RAID1 on a fs with stripe tree.
>>
>> Then the system RAID1 chunks also need stripe tree for zoned devices.
>>
>> This means we're unable to bootstrap at all.
>>
>> Or did you mean, make system chunks RAID1 but not using stripe tree?
>> That can solve the boot strap problem, but it doesn't really look
>> elegant to me...
> 
> RAID1 on zoned only needs a stripe tree for data, not for meta-data/system,
> so it will work and we can bootstrap from it.
> 
That sounds good.

And in that case, we don't need to put stripe tree into system chunks at 
all.

So this method means, stripe tree is only useful for data.
Although it's less elegant, it's much saner.

Thanks,
Qu