Hi Jakub, On 29/05/2026 07:42, Jakub Kicinski wrote: > On Wed, 27 May 2026 22:11:33 +1000 Matthieu Baerts (NGI0) wrote: >> Bufferbloat is baaaad, even in our selftests: let's kill it (or at least >> reduce it). By doing that, the tests (seem to) have a more stable >> transfer, and are then less unstable. That's what patches 1-2 are doing, >> and they can be backported up to 5.10. > > Could you explain a little more what this is actually fixing?
The simult_flows.sh test simulates very low speed links using netem qdiscs. With the current configuration, still a large amount of packets can be queued, which means high TCP RTT samples and then large receive buffer size. Minimal inaccuracy in the pacing rate can lead to using only one subflow towards the end of the connection for a considerable amount of data, while the test expects to use the 2 available paths in parallel. > Does it give a huge increase in test stability? I would say no: the test was not that flaky, mainly on NIPA in fact, and around 4% in the last 100 tests [1]. On my side, I had to get my system busy to reproduce the issues. My hope is that the success rate on NIPA switches from 96 to ~100%. No failure since its introduction in "net-next-2026-05-27--15-00". I sent them to net, just to increase the stability when validating stable kernels as well. But all of this is not urgent, and this series can be applied in net-next if preferred. Do you want me to resend this series targeting net-next? [1] https://netdev.bots.linux.dev/flakes.html?tn-needle=simult-flows-sh Cheers, Matt -- Sponsored by the NGI0 Core fund.
