On 2/1/24 22:25, Dmitry Safonov wrote:
Hi Jakub,
On 2/1/24 21:21, Jakub Kicinski wrote:
On Thu, 1 Feb 2024 00:50:46 +0000 Dmitry Safonov wrote:
Please, let me know if there will be other issues with tcp-ao tests :)
Going to work on tracepoints and some other TCP-AO stuff for net-next.
Since you're being nice and helpful I figured I'll try testing TCP-AO with debug options enabled :) (kernel/configs/debug.config and kernel/configs/x86_debug.config included),
Haha :)
that slows things down and causes a bit of flakiness in unsigned-md5-* tests:
https://netdev.bots.linux.dev/flakes.html?br-cnt=75&tn-needle=tcp-ao
This has links to outputs: https://netdev.bots.linux.dev/contest.html?executor=vmksft-tcp-ao-dbg&pa...
If it's a timing thing - FWIW we started exporting KSFT_MACHINE_SLOW=yes on the slow runners.
I think, I know what happens here:
# ok 8 AO server (AO_REQUIRED): AO client: counter TCPAOGood increased 4 => 6 # ok 9 AO server (AO_REQUIRED): unsigned client # ok 10 AO server (AO_REQUIRED): unsigned client: counter TCPAORequired increased 1 => 2 # not ok 11 AO server (AO_REQUIRED): unsigned client: Counter netns_ao_good was not expected to increase 7 => 8
for each of tests the server listens at a new port, but re-uses the same namespaces+veth. If the node/machine is quite slow, I guess a segment might have been retransmitted and the test that initiated it had already finished. And as result, the per-namespace counters are incremented, which makes the test fail (IOW, the test expects all segments in ns being dropped).
So, I should do one of the options:
- relax per-namespace checks (the per-socket and per-key counters are checked)
- unshare(net) + veth setup for each test
- split the selftest on smaller ones (as they create new net-ns in initialization)
Actually, I think there may be an easier fix:
4. Make sure that client close()s TCP-AO first, making it twsk. And also make sure that net-ns counters read post server's close().
Will do this, let's see if this fixes the flakiness on the netdev bot :)
I'd probably prefer (2), albeit it slows down that slow machine even more, but I don't think creating 2 net-ns + veth pair per each test would add a lot more overhead even on some rpi board. But let's see, maybe I'll just go with (1) as that's really easy.
I'll cook a patch this week.
Thanks, Dmitry