From owner-freebsd-hackers  Fri May 30 09:04:55 1997
Return-Path: <owner-hackers>
Received: (from root@localhost)
          by hub.freebsd.org (8.8.5/8.8.5) id JAA22466
          for hackers-outgoing; Fri, 30 May 1997 09:04:55 -0700 (PDT)
Received: from tornado.cisco.com (tornado.cisco.com [171.69.104.22])
          by hub.freebsd.org (8.8.5/8.8.5) with ESMTP id JAA22461
          for <hackers@freebsd.org>; Fri, 30 May 1997 09:04:53 -0700 (PDT)
Received: from bmcgover-pc.cisco.com (bmcgover-pc.cisco.com [171.69.104.147]) by tornado.cisco.com (8.8.5-Cisco.1/8.6.5) with ESMTP id LAA12933; Fri, 30 May 1997 11:58:40 -0400 (EDT)
Received: from bmcgover-pc.cisco.com (localhost.cisco.com [127.0.0.1])
	by bmcgover-pc.cisco.com (8.8.5/8.8.5) with ESMTP id MAA04390;
	Fri, 30 May 1997 12:04:22 -0400 (EDT)
Message-Id: <199705301604.MAA04390@bmcgover-pc.cisco.com>
To: luiz@nlink.com.br
cc: hackers@freebsd.org
Subject: Cyclades card problems...
Date: Fri, 30 May 1997 12:04:21 -0400
From: Brian McGovern <bmcgover@cisco.com>
Sender: owner-hackers@freebsd.org
X-Loop: FreeBSD.org
Precedence: bulk

The problem is directly collated between interrupt latency and the number
of silo overflows. 

The "problem" is that the Cyclom-Y cards only have DTR flow control
available to them. Note that this means no RTS/CTS flow control that most
applications use for hardware flow control.

Therefore, if the buffers on the card fill up, and the system doesn't service
them in time, a silo overflow will be the result unless XON/XOFF flow
control is enabled.

The PCI card, using a low priority interrupt handler, is the primary cause
for problems on the PCI version (the Ys have, if I remember correctly, as
this conversation was back in December, only 12 bytes of FIFO space).The
system just doesn't get to the card soon enough to avoid a silo overflow.

In the ISA case, the problem is lessened, but an overburdened box will have
the same issues (and I've caused this in a pentium pro 200 running 16 ports
at 38400 at full saturation).

Now, although I can't speak for Cyclades, when I raised the issue with them
back in December (now remember, I work for Cisco, and we're talking port
requirements that will approach 10,000+ between all of our groups), they
didn't seem interested in fixing it. After all, it would require a complete
reengineering of the card.

Currently, all of their energies are being routed towards the Cyclom-Z card,
which is currently available in an 8 port version. The card can do up to 460K
per port on all of the ports. There is a 64 port version due out this summer,
and it'll handle 920+K per second. Depending on customer demand, there
may also be a 128 port version later in the year.

The downside to the Z card is this: There currently isn't a fully working
driver. I have one thats around 80% working. I have a problem with the read()
routine spinning with pppd, its not (properly) generating a HUPCL message
to shells, and the open routine needs to be finished to handle all conditions
that may arise (ie - "allowable" conditions are rejected as EBUSY).

I'm hoping to have something reasonable within a week or two, but thats
where it stands.

If you have any questions, I'll be more than happy to field them.
	-brian