Re: [PATCH v2 0/5] btrfs-progs: check: properly generate csum for various complex combinations

[David Sterba <dsterba@suse.cz>]

On Fri, Jan 14, 2022 at 08:51:18AM +0800, Qu Wenruo wrote:
> [BUG]
> Issue #420 mentions that --init-csum-tree creates extra csum for
> preallocated extents.
> 
> Causing extra errors after --init-csum-tree.
> 
> [CAUSE]
> Csum re-calculation code doesn't take the following factors into
> consideration:
> 
> - NODATASUM inodes
> - Preallocated extents
>   No csum should be calculated for those data extents
> 
> - Partically written preallocated csum
>   Csum should only be created for the referred part
> 
> So currently csum recalculation will just create csum for any data
> extents it found, causing the mentioned errors.
> 
> [FIX]
> - Bug fix for backref iteration
>   Firstly fix a bug in backref code where indirect ref is not properly
>   resolved.
> 
>   This allows us to use iterate_extent_inodes() to make our lives much
>   easier checking the file extents.
> 
> - Code move for --init-csum-tree
>   Move it to mode independent mode-common.[ch]
> 
> - Fix for --init-csum-tree
>   There are some extreme corner cases to consider, in fact we can not
>   really determine a range should have csum or not:
>   * Preallocated, written, then hole punched range
>     Should still have csum for the written (now hole) part.
> 
>   * Preallocated, then hole punched range.
>     Should not have csum for the now hole part.
> 
>   * Regular written extent, then hole punched range
>     Should have csum for the now hole part.
> 
>   But there is still one thing to follow, if a range is preallocated,
>   it should not have csum.
> 
>   So here we go a different route, by always generate csum for the whole
>   extent (as long as it's not belonging to NODATASUM inode), then remove
>   csums for preallocated range.
> 
> - Fix for --init-csum-tree --init-extent-tree
>   The same fix, just into a different context
> 
> - New test case
> 
> [CHANGELOG]
> v2:
> - Handle written extents then hole punched cases
>   Now we always generate csum for a data extent as long as it doesn't
>   belong to a NODATASUM inode, then remove the csum for preallocated
>   range.
> 
> - Enhance test case to include hole punching and compressed extents

Added to devel, thanks, I guess this fixes some of the problems that I
saw when trying to implement the checksum algo switch.