From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Miller Subject: Re: [PATCH] tcp: splice as many packets as possible at once Date: Tue, 13 Jan 2009 16:07:21 -0800 (PST) Message-ID: <20090113.160721.68474279.davem@davemloft.net> References: <20090111133533.GB25337@ioremap.net> <496A17A5.7010200@cosmosbay.com> <20090111160548.GA362@ioremap.net> Mime-Version: 1.0 Content-Type: Text/Plain; charset=us-ascii Content-Transfer-Encoding: 7bit Cc: dada1@cosmosbay.com, w@1wt.eu, ben@zeus.com, jarkao2@gmail.com, mingo@elte.hu, linux-kernel@vger.kernel.org, netdev@vger.kernel.org, jens.axboe@oracle.com To: zbr@ioremap.net Return-path: Received: from 74-93-104-97-Washington.hfc.comcastbusiness.net ([74.93.104.97]:52130 "EHLO sunset.davemloft.net" rhost-flags-OK-FAIL-OK-OK) by vger.kernel.org with ESMTP id S1756718AbZANAHV (ORCPT ); Tue, 13 Jan 2009 19:07:21 -0500 In-Reply-To: <20090111160548.GA362@ioremap.net> Sender: netdev-owner@vger.kernel.org List-ID: From: Evgeniy Polyakov Date: Sun, 11 Jan 2009 19:05:48 +0300 > On Sun, Jan 11, 2009 at 05:00:37PM +0100, Eric Dumazet (dada1@cosmosbay.com) wrote: > > While we drop lock in skb_splice_bits() to prevent the deadlock, we > > also process backlog at this stage. No need to process backlog > > again in the higher level function. > > Yes, but having it earlier allows to receive new skb while processing > already received. > > > > I think that even with non-blocking splice that release_sock/lock_sock > > > is needed, since we are able to do a parallel job: to receive new data > > > (scheduled by early release_sock backlog processing) in bh and to > > > process already received data via splice codepath. > > > Maybe in non-blocking splice mode this is not an issue though, but for > > > the blocking mode this allows to grab more skbs at once in skb_splice_bits. > > > > skb_splice_bits() operates on one skb, you lost me :) > > Exactly, and to have it we earlier release a socket so that it could be > acked and while we copy it or doing anything else, the next one would > received. I think the socket release in skb_splice_bits() (although necessary) just muddies the waters, and whether the extra one done in tcp_splice_read() helps at all is open to debate. That skb_clone() done by skb_splice_bits() pisses me off too, we really ought to fix that. And we also have that data corruption bug to cure too.