Hi, I mentioned in the architecture call last week two examples of dynamic kernel configuration appearing by default with kata-runtime that came somewhat as a surprise to me, in particular compared to what we'd get with a regular container (e.g. crun). Sorry for the delay in sharing details, here they come. -- Examples: - fq_codel instead of noqueue for default virtio-net interface: - crun on Fedora 32: [root@300cd72baa94 /]# ip li sh eth0 3: eth0@if11: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default link/ether de:87:26:08:c3:25 brd ff:ff:ff:ff:ff:ff link-netnsid 0 - kata-runtime: [root@420e660f3870 /]# ip li sh eth0 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel state UP mode DEFAULT group default qlen 1000 link/ether 86:ab:39:73:7c:b8 brd ff:ff:ff:ff:ff:ff This seems to come from runtime/vendor/github.com/vishvananda/netlink/qdisc.go. I don't know the exact reason, fq_codel might be a sane default choice for almost any environment, and I didn't observe any breakage due to this. - nodad on IPv6 addresses: - crun: [root@300cd72baa94 /]# ip ad sh eth0 3: eth0@if11: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default link/ether de:87:26:08:c3:25 brd ff:ff:ff:ff:ff:ff link-netnsid 0 inet 10.88.0.27/16 brd 10.88.255.255 scope global eth0 valid_lft forever preferred_lft forever inet6 fe80::dc87:26ff:fe08:c325/64 scope link valid_lft forever preferred_lft forever - kata-runtime: [root@420e660f3870 /]# ip ad sh eth0 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel state UP group default qlen 1000 link/ether 86:ab:39:73:7c:b8 brd ff:ff:ff:ff:ff:ff inet 10.88.0.28/16 brd 10.88.255.255 scope global eth0 valid_lft forever preferred_lft forever inet6 fe80::84ab:39ff:fe73:7cb8/64 scope link nodad valid_lft forever preferred_lft forever This was introduced by: https://github.com/kata-containers/agent/pull/722/commits/c66b9279cc8ee27397... the reason is clearly explained in comments and was also reiterated by Archana last week, and totally makes sense to me in the general case. I wonder, in this particular case, if we can really assume that any upper stack or any environment can actually guarantee DAD is not needed, but I also couldn't observe any issue due to this, so far. -- My concern is not so much related to networking components themselves at this stage, but rather, more in general, to what might happen with peculiar choices for other subsystems and userspace expectations. I don't have concrete examples about breakages and, again, this is not about build-time kernel configuration. If I understood correctly, bergwolf's proposal from last week goes in the direction of having some kind of facility or approach that allows us to track divergences introduced in kata-runtime compared to regular containers (or to the current configuration of the host kernel), to configure those behaviours and to systematically document particular behaviours. I'm rather new to this, so I don't have a clear picture of how this might apply, in practice, should we choose to start from those two cases as examples. Feedback and further input are warmly welcome! -- Stefano