On 24/10/16 04:05, Phil Pennock wrote:
> On 2016-10-24 at 02:34 +0100, Jeremy Harris wrote:
>> Of course, secondary thoughts, we don't know that the same sequences
>> of library calls on FreeBSD as on Linux will do what's wanted.
> https://people.freebsd.org/~pkelsey/tfo-tools/
server:
rc = setsockopt(listen_fd, IPPROTO_TCP, TCP_FASTOPEN,
&on, sizeof(on));
> * Need to guard against flag being undef
Done for TCP_FASTOPEN, probably need to do for MSG_FASTOPEN.
> * Need to gracefully handle the setsockopt() call failing
OK, so
- The library call sequence is identical
- Use IPPROTO_TCP where Linux has SOL_TCP
- The current coding for failures on the server side is
ok (just ignore the setsockopt failure; we get a
non-fastopen socket as before) for non-supporting kernels
- The current coding for failures on the client side
needs thought. I made it log-noisy but functional
on the grounds that you do an extra syscall
[the sendto(..., MSG_FASTOPEN, ...)] getting a fail
every time on non-supporting kernels. There is
a transport control, defaulting to enabled.
Adding probe code in process startup might help, though
the re-execs we do defeat that. Queue-runners would be
ok but most deliveries are probably foreground.
Should I change the default to not-enabled, for the
medium-term?
--
Cheers,
jeremy
