Proposal: add pru_close protosw method, refactor abort/detach

[gnn at freebsd.org]

At Sun, 18 Jun 2006 02:01:01 +0100 (BST),
rwatson wrote:
> 
> 
> Over the past 6-12 months, I've spent quite a bit of time working with our 
> socket and protocol locking and reference models.  These efforts have been 
> aimed at a couple of things:

Apologies for the length of time I let this sit.

> Attached is a patch that attempts to further rationalize tear-down.
> Specifically, it refactors pru_detach (disconnect and conditionally
> free) and pru_abort (disconnect abruptly and free) into three
> protocol switch functions:
> 
> pru_close: socket has been closed and a sensible shutdown without
> data loss is desired.
> 
> pru_abort: socket is being aborted, generally due to insufficient queue space 
> in a listen socket, or close of a listen socket while connections are waiting 
> to be accepted: close abruptly and potentially with data loss.
> 
> pru_detach: teardown is now unconditional -- both the protocol and socket are 
> done.
> 

This to me is a sensible approach and if it works for both TCP and
SCTP then I suspect it should work for other protocols which are less
complicated.

> The changes require protocol implementors to distinguish close and
> deatch, which while generally clarifying for some protocols (such as
> TCP, where the logic becomes much more clear), for others it has
> never been clear how exactly close worked, and does not become
> clearer.  

I believe your statements above are clear.  Can you put similar,
sensible comments in places where protocols implementors are likely to
see them?  

> What I'm looking for:
> 
> - yay/nay on this approach, and the general change in protosw
>    behavior.  It does touch all protocols, but I think makes things
>    generally more sensible.  this doesn't just support the
>    resock_vertical exploration, but also generally makes things more
>    sensible by moving towards a more typical constructor/destructor
>    and avoiding combining protocol state transitions with socket
>    state freeing.

Yay.

> - Review of the details of the patch.  In particular, help deciding
>    if my splitting up of events between pru_abort, pru_close, and
>    pru_detach for each protocol is right.

It "looks" OK.  Is there a branch you have that I can create a p4
client for and do my usual NetPIPE style abuse testing?  I have been
trying to make a nightly test script for NetPIPE so it might be a good
thing to try on this branch.

Later,
George