last night i turned off distcc support, and got about 200 successful
bootups and zero TCP hangs (as expected - there's not much TCP traffic
if the distcc cluster is not utilized).
but that's 200 overnight tests instead of the expected 600, so this is a
major and rather crippling bug to me.
There's no good way to detect these hung sockets by me from userspace
and get rid of them. Has anyone before thought of the obvious: to write
a kernel-space "TCP socket watchdog" kernel feature that detects them
and tries to free them so that people can become aware of it?
Hung sockets is a re-occuring bug in the TCP stack after all. (and it's
a natural property of it: state machine designs are always vulnerable to
lost event problems.)
Ingo
--
