From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752859AbXCLUVW (ORCPT ); Mon, 12 Mar 2007 16:21:22 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752860AbXCLUVW (ORCPT ); Mon, 12 Mar 2007 16:21:22 -0400 Received: from mga03.intel.com ([143.182.124.21]:19848 "EHLO mga03.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752859AbXCLUVU convert rfc822-to-8bit (ORCPT ); Mon, 12 Mar 2007 16:21:20 -0400 X-ExtLoop1: 1 X-IronPort-AV: i="4.14,275,1170662400"; d="scan'208"; a="194899691:sNHT19374481" X-MimeOLE: Produced By Microsoft Exchange V6.5 Content-class: urn:content-classes:message MIME-Version: 1.0 Content-Type: text/plain; charset="US-ASCII" Content-Transfer-Encoding: 8BIT Subject: RE: [PATCH 1/2] NET: Multiple queue network device support Date: Mon, 12 Mar 2007 13:21:16 -0700 Message-ID: In-Reply-To: <20070312085802.GA1664@ff.dom.local> X-MS-Has-Attach: X-MS-TNEF-Correlator: Thread-Topic: [PATCH 1/2] NET: Multiple queue network device support thread-index: Acdkg/gggoLB5xFYRpaNLuGZAuFArQAXsF0A From: "Waskiewicz Jr, Peter P" To: "Jarek Poplawski" , "Thomas Graf" Cc: "Kok, Auke-jan H" , "David Miller" , "Garzik, Jeff" , , , "Brandeburg, Jesse" , "Kok, Auke" , "Ronciak, John" X-OriginalArrivalTime: 12 Mar 2007 20:21:18.0001 (UTC) FILETIME=[FDB21A10:01C764E3] Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org > -----Original Message----- > From: Jarek Poplawski [mailto:jarkao2@o2.pl] > Sent: Monday, March 12, 2007 1:58 AM > To: Thomas Graf > Cc: Kok, Auke-jan H; David Miller; Garzik, Jeff; > netdev@vger.kernel.org; linux-kernel@vger.kernel.org; > Waskiewicz Jr, Peter P; Brandeburg, Jesse; Kok, Auke; Ronciak, John > Subject: Re: [PATCH 1/2] NET: Multiple queue network device support > > On 09-03-2007 14:40, Thomas Graf wrote: > > * Kok, Auke 2007-02-08 16:09 > >> diff --git a/net/core/dev.c b/net/core/dev.c index > 455d589..42b635c > >> 100644 > >> --- a/net/core/dev.c > >> +++ b/net/core/dev.c > >> @@ -1477,6 +1477,49 @@ gso: > >> skb->tc_verd = SET_TC_AT(skb->tc_verd,AT_EGRESS); > >> #endif > >> if (q->enqueue) { > >> +#ifdef CONFIG_NET_MULTI_QUEUE_DEVICE > >> + int queue_index; > >> + /* If we're a multi-queue device, get a queue > index to lock */ > >> + if (netif_is_multiqueue(dev)) > >> + { > >> + /* Get the queue index and lock it. */ > >> + if (likely(q->ops->map_queue)) { > >> + queue_index = q->ops->map_queue(skb, q); > >> + > spin_lock(&dev->egress_subqueue[queue_index].queue_lock); > >> + rc = q->enqueue(skb, q); > > I'm not sure Dave Miller thought about this place, when he > proposed to save the mapping, but I think this could be not > enough. This place is racy: ->map_queue() is called 2 times > and with some filters (and > policies/actions) results could differ. And of course the > subqueue lock doesn't prevent any filter from a config change > in the meantime. > > After second reading of this patch I have doubts it's the > proper way to solve the problem: there are many subqueues but > we need a top queue (prio here) to mange them, anyway. So, > why not to build this functionality directly into the queue? > There is no difference for a device if skbs are going from > the subqueue or a class, it is only interested in the mapping > result and a possibility to stop and start a subqueue and to > query its status. All this could be done by adding the > callbacks directly to any classful scheduler or, if not > enough, to write some specialized qdisc based on prio. The > possibility to lock only a subqueue instead of a queue could > be analized independently - current proposal doesn't solve > this anyway. > > Regards, > Jarek P. > Thanks again for the feedback. Given some discussions I had last week in the office and the general feedback here, I'm going to remove the new per-queue locking and leave the start/stop functions for each queue and combine entry points for hard_start_xmit(). I'll get this out asap for review once it's been tested here. If we see issues in the future with lock contention on the queues, we can revisit the per-queue locking. Cheers, -PJ Waskiewicz