points
10 months ago
| 3 comments
| HN
Is anyone trying to get those drivers upstreamed?
jabl
10 months ago
[-]
Yes, and yes.

The kernel parts of MOFED are largely backports of the latest and greatest upstream kernel drivers to the various distro kernels their customers actually run. (The non-kernel parts of MOFED is mostly open source but does contain some proprietary special sauce on top, like IIRC SHARP support isn't available in FOSS.). The HPC community does tend to want to use the latest RDMA drivers as those are critical for at scale performance.

For Lustre, the client driver was upstreamed into staging, where it sat AFAIU largely unused for a few years until it was ripped out again. The problem was that Lustre developers didn't adopt an upstream-first development approach, and thus the in-kernel driver was basically a throw over the fence fork that nobody cared about. I think there is an effort to try again and hopefully adopt an upstream-first approach, remains to be seen whether it'll succeed.

reply
pabs3
10 months ago
[-]
For MOFED, why not just wholesale use a newer Linux kernel version?
reply
jabl
10 months ago
[-]
Perhaps the cure is worse than the disease? There are several reasons to stay with a distro kernel:

- Lustre releases target distro kernels, upstream would likely break.

- Distro stays on top of CVE's etc. and provide updates when needed.

- HW likely certified for a few supported distros only, use anything else and you're on your own.

reply
globalc
10 months ago
[-]
That, and if not possible one can try to get the used kABI symbols graylisted at Red Hat, to get informed when they change.
reply
aragilar
10 months ago
[-]
Lustre got dropped from the kernel, so it seems unlikely?
reply