search

andsens / bootstrap-vz

Bootstrap Debian images for virtualized environments
http://bootstrap-vz.readthedocs.io/
Other
263 stars 143 forks source link

Missing ixgbevf ENA drivers for >=Debian stretch #402

Closed joeweoj closed 6 years ago

joeweoj commented 6 years ago

While creating a Debian stretch image my build failed during the ENA setup due to a failure to ungzip a file.

Turns out the drivers file was not a gzipped archive, but was actually an HTML 404 page due to the following missing remote resource https://downloadmirror.intel.com/18700/eng/ixgbevf-4.2.1.tar.gz.

The uri in question is set here https://github.com/andsens/bootstrap-vz/blob/4cb1065551cb905bc97eed08a07718f288a16e82/bootstrapvz/providers/ec2/tasks/network.py#L93

This only affects >= Debian stretch.

To workaround it I hacked my local version to download the file from https://github.com/kwilczynski/packer-templates/blob/master/files/ec2/ixgbevf-4.2.1.tar.gz

Looks like the uri was set when fixing https://github.com/andsens/bootstrap-vz/issues/397

andsens commented 6 years ago

*ugh, this is becoming a real pain. We should find an alternate source for those drivers.

alexfrancavilla commented 6 years ago

Any ideas on fixing it in master? I don't like having a modified master tree on my production pipelines :(

andsens commented 6 years ago

Send a PR and I'll merge :-)

alexfrancavilla commented 6 years ago

@andsens My company would happily provide a mirror (through AWS S3/CloudFront) for these drivers. I would set it up and create a PR then. Are you okay with that in general?

andsens commented 6 years ago

Hm, if you add a shasum check to the download code I don't see a problem.
The drivers are not signed if I recall, so I think others would very much like to not have to trust a third party when it comes to networking code :-)

andsens commented 6 years ago

Also: What about the legality? You should check the licensing regarding distribution before you get yourselves into legal troubles.

alexfrancavilla commented 6 years ago

Adding checksums is of course no problem! I'll check about the legal stuff, thank you for noticing.

justinsb commented 6 years ago

Is there any way to disable certain tasks (other than editing the code)? For example, I don't think we want to install the ixgbevf driver at all on anything with a 4.4 or later kernel - the in-tree version is newer.

I think (but have yet to verify) that we also don't want to install the ENA driver, for the same reason.

andsens commented 6 years ago

Not through the manifest, no. This seems like something we might want to codify though. There's no reason to install a specific driver if we just as well can get it through the kernel.
What Debian version has 4.4 or higher? You can switch on that when assembling the task list.

tnarg commented 6 years ago

@justinsb seems to have a fix https://github.com/justinsb/bootstrap-vz/commit/920fc34cf535b34fade874113852a53119c7c8cb

justinsb commented 6 years ago

That is the updated link, but note that the next PR comments it out entirely :-): https://github.com/justinsb/bootstrap-vz/commit/24800e12e3ef0e09799c8fc742b06b5b1888d2cb

I'm testing this now with stretch, with the in-kernel drivers. So far it seems to work. I previously got confirmation from the intel folk that the in-kernel ixgbevf driver is at least as up to date (despite the different version numbering suggesting that it is old). I don't know the state of the ENA driver, but my understanding is that now that they are in-kernel, the in-kernel versions are maintained just as any other driver would be.

For jessie, I am maintaining an image which includes an updated 4.4 kernel (for running kubernetes https://github.com/kubernetes/kube-deploy/tree/master/imagebuilder), and I was removing the ixgbevf driver manually. Even if we could drop ixgbevf from jessie in favor of the in-kernel version (which I don't think we can), we probably don't want to make that big a change to an existing image - but I don't know the release mechanics here!