udev: add support for AWS EBS NVMe friendly names

[lucab]

Ref: coreos/bugs#2399

Note: while this is AWS-specific, the payload is so small and the lack of updates for OEM partition scares me, so I'm proposing to add it here instead of a new OEM-specific ebuild.

[cgwalters]

I could imagine other cloud providers also doing this; it's not technically specific to AWS right? Or is that hardcoded offset somehow specific to AWS?

I guess my main concern boils down to: This is something we never want to do on a true physical system with real NVMe devices, as there could actually be legacy SCSI devices and we don't want to blow away their device nodes. Also I believe one could (in theory) have NVMe and SCSI devices in libvirt, although in practice people use virtio.

[lucab]

This is very specific to AWS volumes, see coreos/bugs#2399. I'm not aware if/which other providers are doing similar to this. For reference, NVMe devices works perfectly fine without this, it only adds support for user-customized AWS friendly names.

[bgilbert]

I'm okay with shipping this in /usr rather than the OEM.

[lucab]

I manually validated this on a custom AMI on m5.large, both with the rootfs EBS and an hotplugged NVMe volume. I also fixed a minor warning reported by shellcheck. This should be ready to go, PTAL.

[cgwalters]

This is very specific to AWS volumes, see coreos/bugs#2399.

OK. But one thing that isn't clear to me exactly; it looks like we're extracting some sort of raw header from the NVMe device based on a fixed offset; is there any chance that something else is there? Feels like we're at risk of creating random symlinks?

(I could be wrong)

[lucab]

@cgwalters this relies on parsing data out of the vendor-specific payload in the NVMe controller, and vendors can put whatever they want in there. For such reason, the udev rule is scoped to ATTRS{model}=="Amazon Elastic Block Store" as this only applies to AWS as the (virtual) NVMe provider. In order to "read something else" it means that AWS changed their vendor-specific format, which is unlikely as it will break their own compatibility. Even if unlikely that's still possible though, and we'll need to update the parsing/offsets in here in that case, which is another reason for not having this in the OEM partition.

[cgwalters]

For such reason, the udev rule is scoped to ATTRS{model}=="Amazon Elastic Block Store"

Ah, I missed that. That totally addresses my concern, thanks!

[lucab]

@cgwalters on a second look, you are also right because not all the rules are scoped on that. I should fix that, thanks!

[cgwalters]

@cgwalters on a second look, you are also right because not all the rules are scoped on that. I should fix that, thanks!

Maybe simplest to check the model in the script too?

[lucab]

After the last round of discussion, I amended and re-tested this on AWS. udev rules are now all matching on AWS-EBS model, and the script itself also checks that it is being run by udev on an EBS device. PTAL.

[george-angel]

How does one track which version of Container Linux this will make?

[dm0-]

@george-angel It's already released in 1828.0.0, the current alpha.