This tool allows to set TCP options on TCP endpoints via the command line. It can be used, for example, to change the congestion control module or the TCP stack used by TCP endpoints.
It depends on D34138
Details
Details
- Reviewers
glebius lstewart rrs peter.lei_ieee.org rscheff debdrup - Group Reviewers
manpages - Commits
- rG881631a2a371: usr.sbin: add tcpsso
Diff Detail
Diff Detail
- Repository
- rG FreeBSD src repository
- Lint
Lint Skipped - Unit
Tests Skipped
Event Timeline
Comment Actions
When we discussed this tool, weren't generic SOCKOPTs also on the table?
(SO_SNDBUF, SO_DEBUG spring to mind) - the use case would be to modify existing sockets underneath an application, which does not tune these (or when the system defaults can not be changed easily/application restartet).
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 180 | typo? (httpd) | |
| usr.sbin/tcpsso/tcpsso.c | ||
| 156 | not sure, if this scales. OTOH, putting in a parser that interprets tcp.h for all setsockopt stringnames may be overkill... So is the plan to extend these TCP socket options that have some utility over time? | |
Comment Actions
Yes, these are on the table. But I need to figure out if this works with respect to locking. Right now, the socket can't go away, since the socket layer holds a refcount. This is not true when being called via the sysctl interface.
I wanted to bring up things for review early to get some feedback, especially if I'm doing the locking correct.
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 180 | Yepp. Should be thttpd. | |
| usr.sbin/tcpsso/tcpsso.c | ||
| 156 | The plan is to add all socket options which are useful and for which the optval is parsed. Currently I'm only parsing int and char *. Right now, I'm focussing on getting supported what is triggering the development: Change the stack and congestion control module. I added TCP_NODELAY only to have also an example for a socket option which uses the type int for optval. | |
Comment Actions
Only mentioning the first .Nm with a redundant argument, but same applies to all following ones.
Can't check manual page and code are consistent.
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 34 | If I read https://docs.freebsd.org/en/books/fdp-primer/manual-pages/#manual-pages-markup (and specifically 10.3.2. Macros) right, you don't need to specify an argument to .Nm after the first time (in the NAME section). | |
| 63 | ||
| 79 | ||
| 92 | I think you mean .Nm here. | |
| 93 | "apply the option" maybe? | |
| 106 | "apply the option" maybe? | |
| 178–180 | I'd reword this to: To diagnose a problem with a particular TCP connection to | |
| 182 | To match the above suggestion. | |
| 190 | ||
Comment Actions
mandoc -Tlint shows a few nits that I'd like to see corrected before this is committed.
You also have a few lines that exceed 80 characters in width, which I think it'd be good to have reflowed.
Other than that (and the things noted by Pau Amma), mdoc syntax looks good to me.
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 34 | This is correct. | |
| 79 | This is also pointed out by igor. :) | |
Comment Actions
Thanks to the man page reviews. I addressed the issues (the reported ones and the ones reported by mandoc).
Comment Actions
Add support for socket options of level other than IPPROTO_TCP. Also add more symbolic optnames.
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 180 | Well, i assumed this to be a generic stand in for a web demon, rather than a specific one from the ports collection. Perhaps leave this particular reference out from the example completely (its not material to the example itself), just describe it in the text like this. (I assume most people will be familiar with port 80 to be http). If a particular TCP connection to a web server process is behaving in an interesting way, one can determine the .... Or If a particular TCP connection to a web server process ( *thttpd 8* in this example) is behaving in an interesting way, one can determine the .... | |
| usr.sbin/tcpsso/tcpsso.8 | ||
|---|---|---|
| 180 | Switched from thttpd to sshd... This is always installed... | |
Comment Actions
The man page lists only some examples. tcpsso recognizes much more socket option names...
Comment Actions
It's funny because TCP_NODELAY is the first thing people end up enabling (edit: to disable Nagle, he hangs on HN these days as Animats) when they run RPC over a TCP transport on a LAN, the use of Nagle there is itself pretty apocryphal, modulo D-ACK implosion. I ended up doing it on my first "real job" as an A/P Trainee working on Sage Line 50 accounting systems and MS SQL Server in 1999, and ended up doing it when turning up Reuters Dealing 3000 in 2010 in HFT.
Just saying.