Attendees: Claudio Jeker, rscheff@, rrs@, thj@, tuexen@
- Gleb is suggesting to remove the compressed TCP TIMEWAIT state. The memory savings are limited, memory is not that critical anymore and it would simplify the code and allow further optimizations of the code. glebius@ will reach out to people if there are some objections. It was also discussed which information would be need to reduce the time a TCP connections stays in TIMEWAIT state. One thing he will implement is a counter how many compressed states where used at all. Discussion has been started on net@.
ECN implementation has been improved. The code is now contained in one file tcp_ecn.c.
iSCSI changes D32540, D34198, D34230. rscheff@ is looking for reviewers.
D21011 is implementing Accurate ECN and has been updated and will be reviewed by rrs@ and tuexen@.
D23230 is implementing ECN++.
D30043 is a small whitespace cleanup.
D28822 improves PRR, needs some cleanup.
There was bug reported on the current@ mailing list. When a TCP client connects to a server running on the same host, but using a global IPv6 address, the transmission and the first retransmission is reported as received on the interface the global IPv6 address is configured on. The second retransmission is reported on the loopback device. This results in dropping the first two packets if the sysctl variable net.inet6.ip6.source_address_validation is set to 1. The support for this sysctl variable was introduced by this commit. A fix in the IPv6 code is under review D35117. In addition to that, a fix provided by Gleb on the mailing list is needed. This is still being worked on.
- The panic rrs@ was observing is not occuring anymore. There is a code path which should not be taken (and might be relevant for this problem), which should never be executed according to a comment from rwatson@ and the understanding of the code by glebius@, rrs@, and tuexen@. tuexen@ will add a panic() under INVARIANTS to figure out if this code path is actually not taken.
PR 263445 and PR 264257: Both seem to have the same root cause. tuexen@ and rscheff@ are trying to figure out what that root cause is.
glebius@ suggested to retire MSG_EOF from the tree. PR 261699 is a exp-run. tuexen@ will provide some test cases of it at the next meeting to see whether the features works or not. tuexen@ will also test the server side and report.
rrs@ wants to have feedback on D35166. tuexen@ will review it and rrs@ will run some packetdrill scripts.
glebius@ wants feedback on D35199. bz@ isn't fine with it...
- glebius@ has his work on independent socket buffer protocols on review. He suggested to remove the low water marks for datagram sockets. tuexen@ sugested to just fail the setsockopt() calls.
- Claudio Jeker reports some issues the TCP MD5 option and with multipath routing on 13.1. thj@ will help.
Next meeting: 30 June 2022 @ 1500 UTC