On 9/9/21 14:49, David Marchand wrote: > On Fri, Aug 6, 2021 at 2:44 PM Rosemarie O'Riorden <rorio...@redhat.com> > wrote: >> >> If anonymous memory mapping is supported by the kernel, it's better >> to run OVS entirely in memory rather than creating shared data >> structures. OVS doesn't work in multi-process mode, so there is no need >> to litter a filesystem and experience random crashes due to old memory >> chunks stored in re-opened files. >> >> When OVS is not running in memory and crashes, it never reaches the >> clean up scripts that delete the new files it has created, resulting in >> "dirty" memory. OVS will partially overwrite this memory on the next >> start-up, but will fail to run again because its filesystem is full of >> old memory. >> >> Here is an example of these crashes: >> >> https://inbox.dpdk.org/dev/20200910162407.12669-1-david.march...@redhat.com/ > > Overall, I am fine with this change. > > > - In the commitlog, I would focus on the simplication aspect and that > we only use what we need. > Mention of this bug can be removed. > The reason is that this is not as black/white as described: reusing > dirty memory only happened in a special case with containers. > But this does not affect OVS (I don't think we run OVS in containers).
OK. I think, we can just put a period after the 'to litter a filesystem' and remove everything else. That's enough for justification. WDYT? > > > - As mentionned by Ilya, MAP_HUGE_SHIFT is most likely supported on > systems running current OVS. > Can we remove the permissions tweak on /dev/hugepages in > rhel/usr_lib_systemd_system_ovs-vswitchd.service.in ? > (Cc: Timothy who might have an opinion on this). Since users could still put '--legcy-mem' into dpdk-extra, they might still need filesystem access. So, I guess, we still need the tweaks. We could, probably, forbid running with legacy-mem at some point in the future though. Are there plans in DPDK to deprecate the legacy memory model someday? > > > - Somehow related: > * as a followup, we can deprecate --dpdk-hugepage-dir, > * I have been playing with --in-memory and I noticed some bug with the > telemetry socket. > The last started dpdk process takes control of the socket regardless > of other processes, so this happens when running multiple dpdk > applications with none or same XDG_RUNTIME_DIR set in env. > This is more of a fyi: it should not be a problem for people running > ovs with a dedicated user, but it could be a problem for those vilains > that run all their applications as root (/me whistles). > > >> >> Reported at: https://bugzilla.redhat.com/show_bug.cgi?id=1949849 >> Signed-off-by: Rosemarie O'Riorden <rorio...@redhat.com> > > > Thanks, > _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
