Closed legitYosal closed 1 month ago
Ok, here we go,
yesterday I have traced the issue to system calls that were calling repeatedly in systemd, then traced it back to the file descriptor of /proc/*/mountinfo and then boom this file contained 30K lines of repeated mounted points which was not unmounted by docker, then I found the issue on the moby: https://github.com/moby/moby/issues/48305
I am closing this issue cause it is not a bug of OVS
I totally have read this doc: releases and
Open vSwitch userspace is not sensitive to the Linux kernel version. It should build against almost any kernel, certainly against 2.6.32 and later.
But this doc seems a bit old, and I have faced a problem on my stage and production clusters, which was clearly started when tweaking with openvswitch but caused a disturbance in kernel!
I have tested on ubuntu 20 and 22, kernel 5.8 and 6.8, on ovn - ovs 22.03 - 2.17, 22.12 3.0.0 and 24.03 - 3.3.0, and docker latest versions, it seems the issue remains, this is how it will show:
prepend_path
, which can be seen withperf top
there is no space on left device
when after restarting the container it want to unmount containers volumes, it seems it is the cause(?) but it is not, this error can happen after overhead on prepend_path.prepend_path
using systemtap, to thesystemd --user
process(don't see the relation here) (and also a little bit from PID 1)Also have tested this on different versions and still happens, first I was testing for upgrade on stage and then find out on the current version on production we have the same issue, overhead while monitoring perf sometimes goes over 80% percent and it seems
systemd --user
is responsible for calling this function which causes systemd to be congested and unable to start new processes(?), also resources on hypervisor are not used at all!!! and restarting any other container multiple times will not cause such thing.As I have no furthermore clue on what this is and why this is happening, I thought maybe it is because openvswitch compatibility with its kernel module,
@igsilya @ansisatteka