Re: [RFC 0/2] Add dummy RAID stripe tree entries for PREALLOC data

[Josef Bacik <josef@toxicpanda.com>]

On Wed, Sep 18, 2024 at 04:08:48PM +0200, Johannes Thumshirn wrote:
> This is an RFC implementation of Qu's request to be able to
> distinguish preallocated extents in the stripe tree for scrub.
> 
> It's not 100% working yet but only showing the basic "how it's going to
> look like".
> 
> I'm not really sure this is a better solution than returning ENOENT
> and ignoring it in scrub.
> 
> A third possibility would be to do a full backref walk on
> btrfs_map_block() error and then check if it's a preallocated extent.
> 
> Johannes Thumshirn (2):
>   btrfs: introduce dummy RAID stripes for preallocated data
>   btrfs: insert dummy RAID stripe extents for preallocated data
> 

I don't like this approach, I'd rather have a RST represent actually written
bytes on disk rather than adding entries to work around this particular case.

I think that -ENOENT from btrfs_map_block() is liable to result in weird
problems wif we return -ENOENT for other operations.  I think that you change it
to return -ENODATA so that we can for sure trace it back to RST, and use that to
indicate that we need to skip that extent.  This way we have something that is
unique to RST, and we don't have all these other entries that are unrelated to
the RST's job.  Thanks,

Josef