Attendees: jtl, tuexen, thj, Richard
Discussion
- Whitespace commits are in head. tuexen will MFC all relevant changes to stable/12.
Cleanup (combine sack_newdata and snd_recover and reclaim one 32-bit integer): D18811. Michael committed.
Add RFC 6937 support: D18892. This depends on the RFC 6675 support. Richard plans to test more before proposing we commit this. (The testing is also on hold until RFC 6675 support is committed.) Also, needs reviews.
- No update on blackhole detection.
- IP MTU discovery is enabled by default. Discovery works by setting DF bit and lowering MTU when we get back ICMP fragmentation needed responses.
- Blackhole discovery is disabled by default. There is no probe back, so you get stuck at a lower rate permanently.
- jtl@ to provide notes on problems he thought he saw.
- Consensus is to make blackhole discovery "good enough".
Fix RTO timer during SACK retransmissions D23160: Idea is correct; questions about implementation. JTL and Michael will try to look.
- Richard is now having trouble reproducing.
- Richard judges this as perhaps aa less-urgent problem.
- ENOBUF handling in base stack seems suboptimal. RACK/BBR uses a different approach.
Richard reports on a timer problem with Cubic. It doesn't reset the timer when exiting slow start after going idle. Draft patch is D23655, but needs further review.
- Later, we also need to review Cubic and bring it up to the latest RFC.
- jtl added some more people to review.
- ECN:
- tuexen notices (at least) three problems in SYN handling:
- When acting as server and getting IPv6 SYN, the TOS field is not set correctly.
- When retransmitting SYN/ACKs, we are not honoring the max retries sysctl is not set correctly.
- If client resends SYN, we don't update our syncache entry.
- tuexen will submit fixes.
Richard reports an interoperability problem between Linux and FreeBSD when using TCP with ECN. A patch is available at D23364. Michael will review it.
- Extensive discussion on the interaction of reusing snd_recover to track recovery windows for both ECN and CC algorithms. Richard argues from standards that we should have a unified view of the recovery window. tuexen and jtl think that from a code perspective it would be cleaner to have separate tracking to avoid unintended interactions. Richard points out that checking so many references to snd_recover might be difficult. jtl points out that there really aren't that many references. Richard thinks that there are sufficient barriers to prevent us from getting snd_recover use confused.
- rrs@ found some other ECN bugs, and will commit fixes.
- Richard reports Linux and FreeBSD with Cubic and ECN don't play well together due to a (different) bug on each side. Bugs open in each bug tracker.
- tuexen notices (at least) three problems in SYN handling:
Next Meeting
Next meeting is 12 Mar 2020 @ 1400 UTC.