From owner-freebsd-net  Mon Feb  1 10:56:48 1999
Return-Path: <owner-freebsd-net@FreeBSD.ORG>
Received: (from majordom@localhost)
          by hub.freebsd.org (8.8.8/8.8.8) id KAA28380
          for freebsd-net-outgoing; Mon, 1 Feb 1999 10:56:48 -0800 (PST)
          (envelope-from owner-freebsd-net@FreeBSD.ORG)
Received: from pedigree.cs.ubc.ca (pedigree.cs.ubc.ca [142.103.6.50])
          by hub.freebsd.org (8.8.8/8.8.8) with ESMTP id KAA28372
          for <freebsd-net@FreeBSD.ORG>; Mon, 1 Feb 1999 10:56:45 -0800 (PST)
          (envelope-from mjmccut@cs.ubc.ca)
Received: (from ean@localhost) by pedigree.cs.ubc.ca (8.8.8/8.6.9) id KAA21799; Mon, 1 Feb 1999 10:56:44 -0800 (PST)
X400-Received: by /PRMD=ca/ADMD=telecom.canada/C=ca/; Relayed; Mon,  1 Feb 1999 10:56:42 UTC-0800
Date: Mon,  1 Feb 1999 10:56:42 UTC-0800
X400-Originator: mjmccut@cs.ubc.ca
X400-Recipients: non-disclosure:;
X400-Content-Type: P2-1984 (2)
X400-MTS-Identifier: [/PRMD=ca/ADMD=telecom.canada/C=ca/;990201105642]
Content-Identifier: 14826
From: Mark McCutcheon <mjmccut@cs.ubc.ca>
To: freebsd-net@FreeBSD.ORG
Cc: yu.shi@research.nokia.com
In-Reply-To: <199902010642.WAA25830@bubba.whistle.com>
Message-ID: <"14826*mjmccut@cs.ubc.ca"@MHS>
Subject: Re: drop packets?
MIME-Version: 1.0 (Generated by Ean X.400 to MIME gateway)
Sender: owner-freebsd-net@FreeBSD.ORG
Precedence: bulk
X-Loop: FreeBSD.org

Archie Cobbs <archie@whistle.com> writes:

> Shi Yu writes:
>> When the IF output queue is full, the if_output simply drop
>> the packet!??
>
> What else are you going to do with it? :-)

Well, that's simple tail-drop queuing in a FIFO buffer.  It's
certainly the BSD kernel default, but there are arguably more
efficient queuing disciplines for some environments, such as
RED (random early detection) for TCP.  Kenjiro Cho (Sony
Computer Science Lab) has developed an ALTQ (Alternate
Queuing) package for FreeBSD which implements RED, CQB, WFQ,
etc., as well as a template for developing your own queuing
schemes.  See:

  http://www.csl.sony.co.jp/person/kjc/programs.html

>> If the IF output queue is often full, why ip stack does not
>> get feedback from the driver and buffer the packets so that
>> less packets are dropped?
>
> I think ip_output() will return ENOBUFS, if anybody's
> listening.
>
> Anyway, the flow control problem was explicitly left *out* of
> the specification of IP packet delivery. IP packet delivery
> is only "best effort".. and that's exactly what you get.

I think that the intserv and diffserv people would be
disappointed to learn that  ;->

> Higher layer protocols are supposed to deal with flow control
> themselves. For example, TCP.

But many believe that routers need to help with this - hence
RED, RED with ECN, etc.  Also, FRED (fair random early drop)
and similar schemes have been designed to control queue
utilization even by greedy and ill-behaved UDP streams.  So
it's not just tail-drop FIFO anymore - heck, even Cisco's IOS
implements a number of these alternatives....

Mark


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message