From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Hutchings Subject: Re: [RFC] tulip: Support for byte queue limits Date: Tue, 16 Jul 2013 15:37:45 +0100 Message-ID: <1373985465.2120.1.camel@bwh-desktop.uk.level5networks.com> References: <20130712180116.21176.qmail@science.horizon.com> <1373932689.3745.30.camel@bwh-desktop.uk.level5networks.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: George Spelvin , , "Grant Grundler" , "open list:TULIP NETWORK DRI..." To: Grant Grundler Return-path: Received: from webmail.solarflare.com ([12.187.104.25]:63858 "EHLO webmail.solarflare.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932455Ab3GPOhu (ORCPT ); Tue, 16 Jul 2013 10:37:50 -0400 In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On Mon, 2013-07-15 at 18:17 -0700, Grant Grundler wrote: > On Mon, Jul 15, 2013 at 4:58 PM, Ben Hutchings > wrote: > > On Fri, 2013-07-12 at 14:01 -0400, George Spelvin wrote: > >> > Hi George, > >> > While you are right that functionally it doesn't matter, my preference > >> > would be to have nothing between the wmb() and iowrite() that kicks > >> > off the TX. This marginally helps kick off the TX process consistently > >> > slightly sooner. On modern HW, probably irrelevant, but not on the HW > >> > these chips are used on. > >> > >> I'll revise it. It just made sense to me to put it next to the other > >> bookkeeping line of tp->cur_tx++. Should I move them both below the > >> iowrite()? As in: > >> > >> --- a/drivers/net/ethernet/dec/tulip/tulip_core.c > >> +++ b/drivers/net/ethernet/dec/tulip/tulip_core.c > >> @@ -702,11 +702,11 @@ tulip_start_xmit(struct sk_buff *skb, struct net_device *dev) > >> tp->tx_ring[entry].status = cpu_to_le32(DescOwned); > >> wmb(); > >> > >> - tp->cur_tx++; > >> - > >> /* Trigger an immediate transmit demand. */ > >> iowrite32(0, tp->base_addr + CSR1); > >> > >> + tp->cur_tx++; > >> + netdev_sent_queue(dev, skb->len); > > > > This is not good practice, because once you start DMA you have > > effectively passed ownership of the skb to the TX completion handler. > > Is the problem the reference to skb->len? > By passing ownership, are you suggesting the device can change this value? No, the device can complete the descriptor and then the TX completion handler will free the skb. [...] > > But one day someone may want to get rid of this lock, > > so this is a trap waiting to spring. > > Even for tulip driver? Sorry, I just can't imagine anyone taking > enough interest in tulip driver to implement that. I'm not even sure > it would be possible. You're taking interest in it, aren't you? But I accept this is a minor issue. Ben. -- 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.