On Thu, Aug 30, 2007 at 07:49:17PM -0700, [EMAIL PROTECTED] wrote:
<mucho snippage deleted!>
> Whatever it is we do, it is going to have large repercussions
> throughout the source code for the Solaris' IP stack.
Oh yes. And IF we go down this path, we need to examine other
opportunities/problems we can address with a rewhacked message abstraction.
I was recently looking at IPsec performance due to some prodding from
HW-land. We do stuff in ESP today that's pretty awful, part of which is due
to mblk madness:
1.) We have to copy the outbound message because TCP sends a dupb()
block down. We can't encrypt what TCP's storing for possible
retransmissions.
2.) Even if we have the block, we need to make sure we've room to
insert the ESP header into the middle of the packet. Admittedly,
we could do a much better job even with the tools we have today,
but a better messaging abstraction would make the job easier.
(Anyone out there wanna take a crack at the existing stuff? If
so, see RFE #6502521:
http://bugs.opensolaris.org/view_bug.do?bug_id=6502521)
There's early-90s research on this topic out there in the
literature. A search on "pbufs", for example, would yield some
useful data here.
IF (and it's a big IF) we revisit the messaging abstraction we need to put a
lot of thought into what it should do. This includes scanning the existing
research on the subject to make sure we don't repeat mistakes or re-invent
the wheel.
Dan
_______________________________________________
networking-discuss mailing list
[email protected]
