From mboxrd@z Thu Jan 1 00:00:00 1970 Received: (from majordomo@localhost) by oss.sgi.com (8.11.2/8.11.3) id g2RLJHD18176 for linux-mips-outgoing; Wed, 27 Mar 2002 13:19:17 -0800 Received: from smtp.web.de (smtp02.web.de [217.72.192.151]) by oss.sgi.com (8.11.2/8.11.3) with SMTP id g2RLJDq18173 for ; Wed, 27 Mar 2002 13:19:14 -0800 Received: from [80.141.91.231] (helo=there) by smtp.web.de with smtp (WEB.DE(Exim) 4.43 #48) id 16qKqX-00062M-00; Wed, 27 Mar 2002 22:20:41 +0100 Content-Type: text/plain; charset="iso-8859-1" From: Harald Koerfgen Organization: none to speak of To: "Maciej W. Rozycki" , Dave Airlie Subject: Re: [patch] linux: declance multicast filter fixes Date: Wed, 27 Mar 2002 22:20:46 +0100 X-Mailer: KMail [version 1.3.2] Cc: Ralf Baechle , linux-mips@fnet.fr, linux-mips@oss.sgi.com References: In-Reply-To: MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Message-Id: Sender: owner-linux-mips@oss.sgi.com Precedence: bulk Hi, On Monday 25 March 2002 12:40, Maciej W. Rozycki wrote: > Well, the core seems to be already separated -- see drivers/net/7990.c. > I haven't yet checked how suitable it is and many front-end drivers use it > already. > > For the I/O ASIC front-end I'm going to check if the ASIC is capable of > mapping the LANCE more sensibly before starting any further work. The > current configuration is a major loss, doubling the CPU's work and I can't > see any reasonable explanation for such a setup. It's been quite some time since I have hacked the declance driver and I don't remember all the details, so take the following with a grain of salt. The 7990 is basically a 16-bit chip in a 32-bit environment, and, AFAIR, uses two different DMA modes to access host memory. One is a 16-bit word-by-word access for the ring descriptors and the other is 8 16-bit-words-bursts for accessing the ring buffers themselves, where the LANCE only generates one target address per burst. The IOASIC is, just as the CPU, only capable of doing 32-bit transfers to/from memory. So 16-bit LANCE accesses are translated into 32-bit IOASIC accesses but a part of the DMA target addresses are generated by the LANCE. This leads to a 16-bit -> 32-bit mapping for the ring descriptors and a 8*16-bit -> 8*32-bit mapping for ring buffers. Very efficient :-(. Greetings, Harald