Le 05/06/2024 à 20:41, Pablo Neira Ayuso a écrit : > On Wed, Jun 05, 2024 at 11:09:31AM +0200, Nicolas Dichtel wrote: >> Le 05/06/2024 à 10:55, Florian Westphal a écrit : >>> Nicolas Dichtel <nicolas.dichtel@xxxxxxxxx> wrote: >>>> Since the below commit, there are regressions for legacy setups: >>>> 1/ conntracks are created while there are no listener >>>> 2/ a listener starts and dumps all conntracks to get the current state >>>> 3/ conntracks deleted before the listener has started are not advertised >>>> >>>> This is problematic in containers, where conntracks could be created early. >>>> This sysctl is part of unsafe sysctl and could not be changed easily in >>>> some environments. >>>> >>>> Let's switch back to the legacy behavior. >>> >>> :-( >>> >>> Would it be possible to resolve this for containers by setting >>> the container default to 1 if init_net had it changed to 1 at netns >>> creation time? >> >> When we have access to the host, it is possible to allow the configuration of >> this (unsafe) sysctl for the pod. But there are cases where we don't have access >> to the host. >> >> https://docs.openshift.com/container-platform/4.9/nodes/containers/nodes-containers-sysctls.html#nodes-containers-sysctls-unsafe_nodes-containers-using > > conntrack is enabled on-demand by the ruleset these days, such monitor > process could be created _before_ the ruleset is loaded? It's not so easy :) There are several modules in the system. I understand it's "sad" to keep nf_conntrack_events=1, but this change breaks the backward compatibility. A container migrated to a host with a recent kernel is broken. Usually, in the networking stack, sysctl are added to keep the legacy behavior and enable new systems to use "modern" features. There are a lot of examples :)
