autopkg test timeouts on arm64
Bug #2080346 reported by
Matthias Klose
This bug affects 1 person
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
Auto Package Testing |
Fix Released
|
Undecided
|
Skia | ||
rust-reqwest (Ubuntu) |
Fix Committed
|
Undecided
|
Skia |
Bug Description
as seen on https:/
the autopkg tests fail on arm64, always on bos03, the only succeeding one succeeded on bos01.
How should we proceed with that?
Related branches
~hyask/autopkgtest-cloud:skia/iptables_for_rust-reqwest
- Paride Legovini: Approve
-
Diff: 23 lines (+7/-1)1 file modifiedcharms/focal/autopkgtest-cloud-worker/autopkgtest-cloud/worker-config-production/setup-canonical.sh (+7/-1)
Changed in rust-reqwest (Ubuntu): | |
assignee: | nobody → Skia (hyask) |
status: | New → Confirmed |
Changed in auto-package-testing: | |
importance: | Undecided → Critical |
importance: | Critical → Undecided |
Changed in rust-reqwest (Ubuntu): | |
status: | Confirmed → Triaged |
status: | Triaged → Fix Committed |
Changed in auto-package-testing: | |
status: | New → Fix Released |
assignee: | nobody → Paride Legovini (paride) |
assignee: | Paride Legovini (paride) → Skia (hyask) |
To post a comment you must log in.
iirc fron a previous investigation, here are some details on the issue:
The two failing tests are always many_timeout
* connect_
* connect_timeout
If I'm correct, those tests are trying to reach 10.255.255.1, which is supposed to be unavailable and raise a timeout but, maybe due to the proxy (maybe not, I'm not sure), that address is reachable in the infra, making the connection not to timeout, except when something (machine, network, etc...) is slow enough, and that explains why we sometimes see them pass.
One easy way to reproduce the problem locally is to have the 10.255.255.1 IP reachable. many_timeout, that also uses it
In the `rust-reqwest` source package:
```
sudo ip addr add 10.255.255.1/32 dev enp5s0
sudo ip addr add 10.255.255.2/32 dev enp5s0 # only for connect_
cargo test timeout
```
Maybe one possible way forward could be to blackhole those IPs on the testbed with our setup scripts. I'll try to experiment a bit with that solution and see how it goes.