Re: [RFC 00/15] btrfs: RAID5 with RAID stripe-tree (zoned + non-zoned)

On 6/21/26 1:35 AM, Qu Wenruo wrote:
>
> And that in theory removes the need for RAID56 rotation, or the need 
> to have a fixed stripe length.
>
> One of the reason that RAID56 needs rotation is to avoid sub-stripe 
> writes causing too much workload for the P/Q stripe, but since we 
> never over-write in this case, we can even skip the rotation part.
>
> Without rotation, it can further improve partial writes, mentioned below.
[...]
>
> One thing I'm not sure if, do the RAID56 RST still requires the 
> existing 64K stripe len.
>
> E.g. if there is a partial tail that is nr_data * blocksize aligned, 
> but not reaching full nr_data * 64K.
>
> In that case, we can just map that partial write into all devices 
> without extra padding, aka variable stripe length.
[...]
>
> Then limit the RAID56 chunk allocator to always allocate power-of-2 
> number of data devices.
>
> This completely removes the need of partial stripe handling. 
No we can definitively do that. We could also ditch the RAID5 rotation 
(as you said) and go to a RAID4 style parity device (or more for higher 
order parities).