[Bug 2139667] Re: [NEEDS PACKAGING] [mlnx5 hardware not supported] NVIDIA doca-ofed-25.10-dkms DKMS suite

> as this is an extra package to be manually installed by a customer

This isn't an acceptable justification for waving away regression risk
when adding something to the Ubuntu main archive. Any user may install
packages we publish into the main archive whether they are customers or
not, and all users expect us not to regress them.

Please perform a regression analysis again, but this time consider all
users, not just customers.

--
You received this bug notification because you are subscribed to linux
in Ubuntu.
Matching subscriptions: Bgg, Bmail, Nb
https://bugs.launchpad.net/bugs/2139667

Title:
[NEEDS PACKAGING] [mlnx5 hardware not supported] NVIDIA doca-
ofed-25.10-dkms DKMS suite

Status in linux package in Ubuntu:
In Progress
Status in linux source package in Noble:
In Progress

Bug description:
Support for ConnectX hardware through doca-ofed-25.10 DKMS in Ubuntu
Noble

[ SRU Justification ]

This is an effort to repackage the DKMS portion of the doca-ofed upstream package distributed by NVIDIA through a public HTTP source (e.g. https://linux.mellanox.com/public/repo/doca/).
NVIDIA distributes this as a set of deb packages in a tarball at the following address: https://developer.nvidia.com/doca-downloads; this package is for the 3.2.1 version of the upstream NVIDIA package.

Current Ubuntu packaging lacks support for NVIDIA Bluefield/IGX/DGX
platform: this is needed as customers using NVIDIA hardware can take
advantage of custom, high performance drivers directly through
Canonical apt system, without the need of downloading an external
tarball to be manually installed on the target machines.

This helps as well any automated installation flow as, by just
specifying the need for this package as an apt dependency, avoids the
automation issue of finding the HTTP address of the latest version of
the doca-ofed package distributed by NVIDIA and the relative scripting
to manually install them.

This package adds a set of kernel modules related to the NVIDIA mlnx5
hardware family (e.g. IB-capable ConnectX devices).

This package is related to the https://launchpad.net/ubuntu/+source/mofed-modules-24.10/ package already accepted last year. This is the newer LTS version distributed by NVIDIA.
There are no modification in repackaging, apart from a more coherent name with NVIDIA naming scheme, has been applied to the repackaging.
Another difference from the 24.10 version, is a smaller list of .ko files that will be built on the target machine, following what invoking the doca-ofed target on the original package installs.

[ Current Test Plan ]

 * Hardware tested on:
     * Noble-nvidia-tegra noble-derivative kernel 6.8.0-1013.13 -> 6.8.0-1016.16
         - Tested on IGX Developer tool - ARM64
     * Noble GA kernel 6.8.0-101-generic
         - Tested on DGX B200 (amd64) and IGX GH200 (arm64)

  * Compile and module insertion test on:
     * Noble GA 6.8.0-71-generic, -88-generic, -90-generic and -100-generic
     * Noble HWE-6.14.0-37-generic
     * Noble HWE-6.17.0-14-generic
     * Questing 6.17.0-7-generic and 6.17.0-8-generic

 * Before uploading to Ubuntu an update version of the repackaged
kernel modules, a local test to the affected kernels must be executed.
At the moment, the main target kernel is the noble:linux-nvidia-tegra
one, so the tests will be thoroughly executed on the specific hardware
running said kernel. If the modules load correctly, the package is
considered sane and can be uploaded to the wider public. The
compilation test is executed as well on the latest version of Noble GA
and HWE.

* Another test suite run on each new spin of the module, is a smoke test running on DGX B200 (amd64) and GH200 (arm64) machines; those machines contains the hardware this set of kernel modules target.
The suite is able to test for the produced modules to be able to run IB interactions, fully load all the built modules and interactions between kernel-space modules/user-space applications.
A full test run can be found on the ARGOS-1951 Jira ticket.
Test-plan can be found at https://github.com/canonical/mofed-userspace-integration/tree/main/testcases
Test machines are freely available in Testflinger.

* The produced modules has been tested by an Nvidia team that has
given us the green light for the release.

[ Where problems could occur ]

 * No issues are expected, as this is an extra package to be manually
installed by a customer and, without the necessary hardware, no driver
is automatically loaded in memory.

[ Other Info ]

 * The repackaging code is contained in the repository at the
following address:
https://kernel.ubuntu.com/forgejo/alessiofaina/nvidia-doca-ofed-dkms

 * Examples of the repackaged DKMS can be found at the following
personal PPA:
https://launchpad.net/~alessiofaina/+archive/ubuntu/mofed-autoupload-
test/+packages

* All the modules as GPL-2 or Dual BSD/GPL licensed, no closed source
binaries are redistributed.

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2139667/+subscriptions