Re: Changing the checksum algorithm on an existing BTRFS filesystem

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2021/10/5 00:01, Matthew Warren wrote:
> I don't know how btrfs is layed out internally, but is checksum tree
> separate from file (meta)data or is it part of the (meta)data? If it's
> separate it should be possible to build a second csum tree and then
> replace the old one once it's done, right?

There are several problems, even for off-line covert.

1) Metadata checksum
    Unlike data checksum, metadata checksum is inlined into the metadata
    header.
    Thus there is no way to make the csum co-exist for metadata, and we
    need to re-write (CoW) all metadata of the fs to convert them to the
    new algo.

2) Superblock flags/csum_type
    Currently we have only one csum_type, which is shared by both data
    and metadata.

    We need extra csum_type to indicate the destination csum algo.
    We will also need to sync this bit to kernel, no matter if the kernel
    really uses that.

Thanks,
Qu

>
> Matthew Warren
>
> On Mon, Oct 4, 2021 at 4:52 AM David Sterba <dsterba@suse.cz> wrote:
>>
>> On Mon, Oct 04, 2021 at 12:26:16AM -0500, Matthew Warren wrote:
>>> Is there currently any way to change the checksum used by btrfs
>>> without recreating the filesystem and copying data to the new fs?
>>
>> I have a WIP patch but it's buggy. It works on small filesystems but
>> when I tried it on TB-sized images it blew up somewhere. Also the WIP is
>> not too smart, it deletes the whole checksum tree and rebuilds if from
>> the on-disk data, so it lacks the verification against the existing
>> csums. I intend to debug it some day but it's a nice to have feature,
>> I'm aware that people have been asking for it but at this point it would
>> be to dangerous to provide even the prototype.