On Wed, May 6, 2020 at 5:37 AM Peng Tao <tao.peng@linux.alibaba.com> wrote:
On 2020/5/5 21:27, Stefano Brivio wrote:
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. Hi Stefano,
My main concern about making guest kernel behave like the host kernel is that we might lose the ability to have a customized/optimized kernel just for container use case. There are a lot of kernel config options that are not going to be useful for container workload. So instead of just using the host kernel (for kata containers), I would suggest just using a minimal guest kernel as a basis and start adding new config options/modules as we identify new needs. And that is what we have been doing for Kata Containers in the past years.
Production wise there is a lot of value in having the same kernel on the host and the guest. For example, taking a workload that has been run as a vanila container and then running it on a kata container could require a testing/certification process from scratch if the host/guest kernels are different. Kernel CVEs would also be better managed if the host/guest kernels are the same.
As far as dynamic kernel options, is it enough to use kernel boot options + sysctls + kernel modules to meet the requirement? I asked it in the AC meeting but there was no conclusion yet. Most containers shouldn't care about specific kernel options. If kernel boot options + sysctls + kernel modules still cannot satisfy it, that should be a very specialized workload and IMO such users (who would likely be very professional at customizing things) can use a customized guest kernel instead of the standard one shipped with Kata Containers.
Just my 2 cents...
Cheers, Tao
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!
-- Into something rich and strange.
_______________________________________________ kata-dev mailing list kata-dev@lists.katacontainers.io http://lists.katacontainers.io/cgi-bin/mailman/listinfo/kata-dev