From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Hutchings Subject: Re: 8139cp: set ring address before enabling receiver Date: Wed, 21 Nov 2012 20:18:55 +0000 Message-ID: <1353529135.2619.36.camel@bwh-desktop.uk.solarflarecom.com> References: <20120602235020.2C0A57C006C@ra.kernel.org> <1353517042.26346.130.camel@shinybook.infradead.org> <50AD1972.5080403@pobox.com> <1353527481.26346.150.camel@shinybook.infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Jeff Garzik , Jason Wang , "David S. Miller" , To: David Woodhouse Return-path: Received: from webmail.solarflare.com ([12.187.104.25]:34891 "EHLO webmail.solarflare.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756286Ab2KUUS7 (ORCPT ); Wed, 21 Nov 2012 15:18:59 -0500 In-Reply-To: <1353527481.26346.150.camel@shinybook.infradead.org> Sender: netdev-owner@vger.kernel.org List-ID: On Wed, 2012-11-21 at 19:51 +0000, David Woodhouse wrote: > On Wed, 2012-11-21 at 13:12 -0500, Jeff Garzik wrote: > >=20 > > What sticks out at me from the commit message? > >=20 > > It was not tested on the famously quirky 8139 hardware at all. > >=20 > > While I have not looked at the 8139C+ data sheet in a while, someti= mes=20 > > the hardware _did_ have a strange init order. > >=20 > > As this works in a simulator but fails on real hardware, it seems l= ike=20 > > an obvious regression caused by an untested [on read hardware] patc= h. >=20 > The data sheet (v1.6, from http://realtek.info/pdf/rtl8139cp.pdf ) sa= ys > in =C2=A76.33 (C+ Command Register): > "Enable C+ mode functions in C+CR register first, > =3D> Enable transmit/receive in Command register (offset 37h), > =3D> Configure other related registers (ex. Descriptor start address= , > TCR, RCR, ...)." > > I understand the concern expressed in the offending commit message ab= out > DMA happening to invalid addresses, and I'll look at the data sheet > harder to see when the DMA actually starts happening. But it definite= ly > seems that our current code isn't doing what the data sheet says. >=20 > I wonder if I can find one of these lying around and stick it in a > machine with an IOMMU... You might be able to avoid disaster by doing: 1. Set MAC filter to drop everything 2. Enable RX DMA 3. Set RX DMA ring address 4. Set MAC filter according to current flags & multicast list I'm assuming, knowing nothing about this particular hardware, that the MAC filter register(s) will accept writes before RX DMA is enabled. Ben. --=20 Ben Hutchings, Staff Engineer, Solarflare Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.