Skip to contentSkip to conversation

net/udpgro.sh is flaky on debug

TEST Jakub Kicinski 2 participants 3 messages Jul 10, 2025

Subsystems (3)
lore
Share & tools
lore
maintainer
#1 lore raw
Quoted 7 lines · Jakub Kicinski

net/udpgro.sh has around 6% false positive rate on the debug kernels :(

Please see:
https://netdev.bots.linux.dev/contest.html?executor=vmksft-net-dbg&test=udpgro-sh&ld_cnt=200&pass=0
for recent failures.

Is there anything we can do to make it more resilient ?

As most failures happen while waiting for the 'following' packets in non GRO scenarios, a trivial attempt could increasing the inter-packet timeout (currently 10ms).

Something alike the following. I can send a formal patch, and we can let it stage in PW for a bit to observe if it actually helps.

/P
---

(quoted diff) +0 -0

b/tools/testing/selftests/net/udpgro.sh
index 1dc337c709f8..43fcd4b0f6a2 100755

tools/testing/selftests/net/udpgro.sh +0 -1
@@ -48,7 +48,7 @@ run_one() {
 
        cfg_veth
 
-       ip netns exec "${PEER_NS}" ./udpgso_bench_rx -C 1000 -R 10

${rx_args} &
+ ip netns exec "${PEER_NS}" ./udpgso_bench_rx -C 1000 -R 100
${rx_args} &

        local PID1=$!

        wait_local_port_listen ${PEER_NS} 8000 udp
author maintainer
#2 lore raw
Quoted 2 lines · On Thu, 10 Jul 2025 16:20:28 +0200 Paolo Abeni wrote:

Something alike the following. I can send a formal patch, and we can let it stage in PW for a bit to observe if it actually helps.

Added to the local hacks, should start showing up from net-next-2025-07-10--18-00 onward. Thanks!