This is an old revision of the document!
TCP Projects for FreeBSD 8.0
This wiki page is a catalog of known major TCP projects underway for FreeBSD 8.0. More detailed information on the projects may be found below.
| Project | Flag Wavers | Schedule | MFC 7? | Link | Summary |
| TCP Offload | KipMacy | Merge to CVS in progress | desired | ? | Support for complete TCP connection offload to 10gbps network device drivers. |
| TCP overhaul | AndreOppermann | ? | no | ? | Andre is planning a general cleanup of the TCP implementation. |
| TCP pluggable congestion control | SamLeffler, RuiPaulo, AndreOppermann, KipMacy, James Healy, LawrenceStewart | Prototype in svn | desired | tcp_cc svn project | Abstract the congestion control mechanism so that they may be plugged; provide alternative algorithms, and support algorithms present for TOE cards. |
| TCP locking granularity/parallelism | RobertWatson | Planning phase | partially | ? | Improve the granularity of TCP locking, and create new opportunities for parallel execution. |
| TCP ABI cleanup | RobertWatson | Planning phase | desired | tcpabi P4 project | Clean up data structures exported to userspace by TCP sysctls. |
| TCP ECN | RuiPaulo | Testing phase | ? | tcpecn P4 project | TCP Explicit Congestion Notification (ECN) |
| TCP ABC | LawrenceStewart, RuiPaulo | Planning phase | desired | tcp_abc svn project | TCP Appropriate Byte Counting |
| TCP UTO | RuiPaulo (?) | Planning phase | desired | ? | Per-connection timeout period defined by the user. |
Schedule
So far, only a guess at the 8.0 schedule is possible:
| Date | Significance |
|---|---|
| December 2008 | All major TCP changes for 8.0 must be completed to allow six months of testing time before 8.0-RELEASE. |
| June 2009 | Possible 8.0 release date |
Projects
TCP Offload
Support for complete TCP connection offload to 10gbps network device drivers. Kip is the author of the Chelsio 10gbps driver for FreeBSD, and has updated the driver to support the prototype TOE interface now in CVS.
TCP overhaul
TCP pluggable congestion control
Several plans exist to merge support for pluggable congestion control for TCP, as well as provide support for algorithm selection for TOE device drivers. This may require moderate restructuring of TCP to abstract congestion control decisions currently scattered implicitly around the TCP implementation.
TCP locking granularity/parallelism
The TCP implementation is MPSAFE, but the tcbinfo lock protects all global data structures, so is acquired in the input path during lookup and in timers, leading to reduced parallelism. This task works to break down the tcbinfo lock, improving locking granularity, but also attempts to improve parallelism by distributing work over multiple CPUs and introducing weak CPU affinities for TCP connections.
TCP ABI cleanup
Perforce: //depot/user/rwatson/tcpabi/…
The “guts” of TCP data structures are currently exposed to userspace via sysctl, meaning that even minor kernel data structure layouts disturb the monitoring API/ABI used by user processes to track network connections. I.e., netstat. This task creates new exported data structures that contain no kernel-specific structures, such as mutexes or callout structures, and attempts to stabilize monitoring ABIs by adding plenty of padding for future extension.
TCP ECN
Perforce: //depot/projects/tcpecn/…
TCP ECN is a small modification to TCP end hosts and routers that notifies the peers when congestion is about to occur, allowing them to reduce the sending rate.
TCP ABC
TCP appropriate byte counting (RFC 3465) addresses a congestion control (CC) related issue introduced by the early TCP specs. It suggests increasing the congestion window by the number of bytes ACK'ed by a TCP ACK, as opposed to the current scheme which relies on an approximation driven by ACK clocking. ABC will also benefit the modular CC work as many of the new CC algorithms break some of the implicit assumptions required for the currently used approximation to function correctly.
TCP UTO
From the Abstract:
The TCP user timeout controls how long transmitted data may remain unacknowledged before a connection is forcefully closed. It is a local, per-connection parameter. This document specifies a new TCP option - the TCP User Timeout Option - that allows one end of a TCP connection to advertise its current user timeout value. This information provides advice to the other end of the TCP connection to adapt its user timeout accordingly. Increasing the user timeouts on both ends of a TCP connection allows it to survive extended periods without end-to-end connectivity. Decreasing the user timeouts allows busy servers to explicitly notify their clients that they will maintain the connection state only for a short time without