Attendees: glebius@, peterlei, rrs@, rscheff@, rstone@, tuexen@
Unloading CC requires specific lock orders and pushes constraints on the fast path. rrs@ adds using reference counting for removing an CC module in D33249 and tuexen@ will write a userland application to switch CC modules on TCP connections. This tool should be able to modify on a per TCP connection grain
- the CC module being used
- the TCP stack being used
- turn on and off BB logging
- set the send and receive window
- 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@.
D33854 adds tests for DSCP handling in TCP. tuexen@ will add the needed functionality to do this with packetdrill.
rstone@ asks for feedback on two LRO related reviews: D33774 and D33775. rrs@ will review.
Any feedback on 260393? tuexen@ will use packetdrill to test some scenarios suggested by rscheff@. glebius@ mentioned that the kernel dump does not contain critical application data...
Any feedback on the following BBR related syzkaller-issue (no reproducer)? glebius@ will improve the panic() message to include the errno. BBR seems to fail some tcp_output() calls on purpose.
- The handling of the PUSH flag in LRO was discussed. Also the impacts of LRO to ECN and accurate ECN was discussed. rstone@ will put a patch up on phabricator.
- rstone@ mentioned his unit test framework for kernel functions and will continue his work to get this into FreeBSD.
Next meeting: 27 January 2022 @ 1500 UTC