cvs commit: src/sys/netinet tcp_subr.c tcp_var.h

[Mike Silbersack]

On Mon, 19 Apr 2004, Mike Silbersack wrote:

>   Enhance our RFC1948 implementation to perform better in some pathlogical
>   TIME_WAIT recycling cases I was able to generate with http testing tools.
>
>   Except in such contrived benchmarking situations, this problem should never
>   come up in normal usage...  until networks get faster.

I think that we may have to break away from standard RFC handling and
change the TIME_WAIT code in tcp_input so that it will accept any SYN
packet coming in without regard to the sequence number, forcing the
TIME_WAIT socket to be recycled.

The problem is that right now we're using an increment rate of 1MB per
second.  This means that if a connection on a certain port is established
on a gigabit network, more than 1MB/second is sent, the connection is
closed, and then a new connection is attempted with the same port pair, it
will fail because the SYN's sequence number will be less than the sequence
number the previous connection ended with.

We could somewhat address this by increasing the 1MB/second rate to
something greater, but that would increase the rate of serial number
wraparound, which probably isn't a good thing.  Also, with a 32-bit
sequence space, we just can't increase the rate so that it's fast enough
to keep up with 1Gb/sec or 10Gb/second networks.

So, I may tackle this at some point in the future, after I survey the
other OSes and see how they're handling it.

Mike "Silby" Silbersack