LWN: Comments on "The end of modversions?"

The end of modversions?

ndesaulniers — Mon, 26 Aug 2019 23:02:00 +0000

Looks like parsing of attributes on aggregate definitions doesn't work quite right either: https://lkml.org/lkml/2019/8/26/898.

The end of modversions?

lsl — Wed, 29 Nov 2017 19:42:12 +0000

Ideally, those enterprise users would test recent kernels for regressions in features they depend on. When the relevant changes have trickled down to enterprise distribution kernels it's generally too late to do anything about it.

Reading LWN is a good idea, too, obviously.

The end of modversions?

pabs — Wed, 29 Nov 2017 09:25:40 +0000

Getting your code upstream seems like the best option to me.

The end of modversions?

cjmather — Tue, 28 Nov 2017 14:58:00 +0000

The product I develop makes use of this feature. It's particularly useful on Ubuntu which rolls new kernels frequently, but breaks compatibility (for our module) infrequently.

We could try to get our code upstream, but the code is specific to our product and may not be of general interest. We could use DKMS, but it seems brittle to require enterprise customers to have a functioning toolchain just to use our product.

In the bad old days, Unix OS's like HP-UX would guarantee compatibility for kernel extensions across release families (e.g., minor releases). This made it easier for application developers and (ultimately) enterprise customers. I hope that folks keep enterprise customers in mind -- and not testers, as cited in the article -- when evaluating the utility of this feature.

The end of modversions?

JanC_ — Fri, 02 Dec 2016 01:14:16 +0000

I know the Ubuntu kernel team uses a script ('abi-check') that checks for ABI-changes at build-time (used for kernel/package versioning), but I'm not sure if it depends on "modversions" or not…

The end of modversions?

jcm — Thu, 01 Dec 2016 10:11:42 +0000

Oh, and technically, it gets far worse if you rely upon manually maintained ABI revisions. Now you have a situation in which - unless everyone is very careful to always remember to update the versioning whenever anything that depends upon a symbol changes (and that's not just the interface, that's any recursively using a structure referenced from a structure in the interface) - you can have modules that "load" and will crash the kernel. That's exactly the inverse of ideal. The reason for the complex CRC trickery is to ensure that - to a first approximation - nothing in the chain of dependencies in terms of structures used by a symbol has changed.

The end of modversions?

jcm — Thu, 01 Dec 2016 10:06:57 +0000

A number of the Enterprise distributions use modversions as part of kernel ABI. It would be extremely unfortunate to see this feature go away. I've asked a few folks to look into making the case to improve the code, rather than remove it.