xapi-project / xen-api / 41804245234
79%

DEFAULT BRANCH: master
Ran
Jobs 1
Files 36
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 78.996%. Remained the same
41804245234

push

github

web-flow 
CA-432654: Make networkd start lldpd quickly and reliably (#7354)

When Dom0's host name doesn't resolve, lldpd crashes after systemd
starts it on dom0. This is fixed by
https://github.com/lldpd/lldpd/commit/a2f4b2c92.
However it exposes the fragility that xcp-networkd could take more than
45 s to start. systemd then killed it, and networkd sometimes dumped
core on the way down.

There are three causes:
- Config replay race. lldpd reports it is ready before it replays
/etc/lldpd.d. Settings in that file could overwrite what networkd had
just set, so networkd had to wait for the replay to finish. With a large
amount of NICs and lldpd crash, the time is wasted here.
- Slow lookups. Until the system name is set, lldpd looks up the host
name on every update, and every lldpcli config command triggers an
update. With a name that doesn't resolve, each lookup takes about 0.5 s.
networkd sent the system name late, after several other commands, and it
also waited for lldpd's own config replay to finish. Startup slowed down
with every NIC.
- Shutdown deadlock. The SIGTERM handler can run on the startup thread
while that thread holds the LLDP scope lock. Lldp.stop then took the
same lock, failed with "Resource deadlock avoided", and never reached
exit. systemd aborted networkd and it dumped core.

Changes
- Start lldpd with a deny-all interface pattern.
- lldpd is started with -I !* through a runtime systemd drop-in
(/run/systemd/system/lldpd.service.d/10-networkd.conf), so it manages no
interface until networkd sets the scope.
- The default config file now holds only the settings networkd doesn't
manage, so lldpd's replay and networkd no longer overlap. The wait for
the replay is removed.
- A port is added to the scope only after its settings are applied,
because a port in the scope starts sending at once.
- Set the system name first.
- Right after starting lldpd, networkd sets the system name before any
other config, so the slow h... (continued)

3667 of 4642 relevant lines covered (79.0%)

0.79 hits per line

Jobs
ID Job ID Ran Files Coverage
1 python3.11 - 41804245234.1 36
79.0
GitHub Action Run
Source Files on build 41804245234