I hit another race in the TCP connect code related to implied
connects on sendmsg().
I am going to be changing the so_mport stuff a bit to improve code
flow and fix the issue as well as other issues where sockets have
to migrate between cpus. This will also improve performance as the
target message port will be cached in the socket structure. It will
also allow us to do things like spread unix domain sockets and
possibly even socketpairs across multiple cpus.
I will be storing the target port in the socket structure and will
modify the message service loop to forward messages that wind up on
the wrong port (due to the protocol changing the port during e.g.
a connect()).
This will remove a ton of special cases. One special case will be
added to synchronize port changes. For example, if a ton of sendmsg()
calls are made (and the first was for an implied connect), then when
the port is changed the socket will have to ensure that no new
messages are queued to the new port until all messages sitting on the
old port have been properly forwarded.
-Matt
Matthew Dillon
<dil...@backplane.com>
