From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Hutchings Subject: Re: [PATCH] xen network backend driver Date: Wed, 19 Jan 2011 19:25:47 +0000 Message-ID: <1295465147.11126.57.camel@bwh-desktop> References: <1295449318.14981.3484.camel@zakaz.uk.xensource.com> <1295455216.11126.39.camel@bwh-desktop> <1295459316.14981.3727.camel@zakaz.uk.xensource.com> <1295460304.11126.53.camel@bwh-desktop> <4D3738AB.60701@goop.org> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: Ian Campbell , "netdev@vger.kernel.org" , xen-devel , Konrad Rzeszutek Wilk , Tom Herbert To: Jeremy Fitzhardinge Return-path: Received: from exchange.solarflare.com ([216.237.3.220]:7537 "EHLO exchange.solarflare.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754832Ab1ASTZv (ORCPT ); Wed, 19 Jan 2011 14:25:51 -0500 In-Reply-To: <4D3738AB.60701@goop.org> Sender: netdev-owner@vger.kernel.org List-ID: On Wed, 2011-01-19 at 11:16 -0800, Jeremy Fitzhardinge wrote: > On 01/19/2011 10:05 AM, Ben Hutchings wrote: > > Not in itself. NAPI polling will run on the same CPU which scheduled it > > (so wherever the IRQ was initially handled). If the protocol used > > between netfront and netback doesn't support RSS then RPS > > can be used to spread the RX work > > across CPUs. > > There's only one irq per netback which is bound to one (V)CPU at a > time. I guess we could extend it to have multiple irqs per netback and > some way of distributing packet flows over them, but that would only > really make sense if there's a single interface with much more traffic > than the others; otherwise the interrupts should be fairly well > distributed (assuming that the different netback irqs are routed to > different cpus). > > Also, I assume that if most of the packets are not terminating in dom0 > itself but are sent out some other device (either real hardware or to > another domain), then there won't be any protocol processing and the > amount of CPU required to handle the packet is minimal. Is that true? > And if so, would RPS help in that case? I would expect the cost of an > IPI to swamp anything else that needs to happen to the packet. IPIs are apparently pretty cheap now. If NAPI doesn't work for netback, with or without RPS, maybe you have to use that kthread. But please don't reinvent the wheel by creating your own tasklets or require users to set a special parameter. Ben. -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.