Integrating dm-integrity into mdadm/MD.

[ncroxon]

I am thinking about adding support to call dm-integrity to mdadm/md, as a command line parameter.

[mwilck]

Please explain why you think mdadm should support this. What does mdadm have to do with dm-integrity?

[ncroxon]

The Risk: MD RAID does not know which disk is "telling the truth." If a bit flipped on a data disk, MD RAID will simply recalculate the parity disk to match that corrupted bit, effectively "baking in" the corruption.
To truly protect against silent bit flips, it should use a dm-integrity on top of MD RAID’s block devices. This block layer stores checksums for every block. If a bit flips, they can compare the file against its checksum, realize the data is wrong, and use the RAID parity to reconstruct the correct data. While RAID provides high availability, it is not a standalone solution for data integrity. MD RAID should be paired with dm-integrity to ensure that "reconstruction" doesn't inadvertently "bake in" silent corruption.

[mwilck]

I am not sure if I understand. Are you proposing a new technology that combines RAID and verity to achieve a cryptographically sound RAID of sorts?

That would require functionality in the kernel, I suppose? Like: the reconstruction thread in the kernel needs to call back into dm-integrity to figure out in which direction to fix a certain inconsistency? Does this kernel functionlity exist, or are you or is anyone else working on it?

[mwilck]

OK, thanks. So it's meant for this specific setup. I understand. How would your envisaged command syntax look like? AFAIK you need a hash device for every data device. I suppose that currently the user needs to pick these hash devices manually using veritysetup, then call mdadm to assemble the array. How would a user write a create command that automates this? What other commands except "create" do you plan to support?

[ncroxon]

some parameters that can be passed to dm-integrity are:
When formatting:
--integrity (sha1 or sha256 or sha512)
--sector-size 4096
--tag-size 8
And we will need parameter for a name for each formatted device to open.
integritysetup open /dev/sda1 int1 (name). This will create a block device in the /dev/mapper directory named int1.
Which will be used by mdadm as a member device to build the MD array on.
(current method) mdadm -C /dev/md0 -n 4 -l 6 --chunk 2048 --assume-clean /dev/mapper/int[1234]

maybe something like this:
mdadm -C /dev/md0 -n 4 -l 6 --chunk 2048 --integrity sha1 --sector-size 4096 --tag-size 8 /dev/mapper/int[1234]

[mwilck]

You still need integritysetup to format the intX devices?

[ncroxon]

yes, mdadm will shell call it.

[mbroz]

Well, if you want to integrate it (and I guess also to avoid any complex things in dm-integrity configuration like encryption), then IMO you should use libcryptsetup C API directly, not wrapping integritysetup tool.
And if it is not possible, I would like to know why, as we could fix it :)