From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jarek Poplawski Subject: Re: [PATCH net-next-2.6] net: pskb_expand_head() optimization Date: Sun, 12 Sep 2010 22:45:52 +0200 Message-ID: <20100912204552.GA2585@del.dom.local> References: <20100911123140.GA1939@del.dom.local> <20100911.203002.193707453.davem@davemloft.net> <20100912104534.GA2056@del.dom.local> <20100912.085833.226777368.davem@davemloft.net> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: eric.dumazet@gmail.com, netdev@vger.kernel.org To: David Miller Return-path: Received: from mail-wy0-f174.google.com ([74.125.82.174]:64943 "EHLO mail-wy0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753501Ab0ILUp7 (ORCPT ); Sun, 12 Sep 2010 16:45:59 -0400 Received: by wyf22 with SMTP id 22so5390789wyf.19 for ; Sun, 12 Sep 2010 13:45:58 -0700 (PDT) Content-Disposition: inline In-Reply-To: <20100912.085833.226777368.davem@davemloft.net> Sender: netdev-owner@vger.kernel.org List-ID: On Sun, Sep 12, 2010 at 08:58:33AM -0700, David Miller wrote: > From: Jarek Poplawski > Date: Sun, 12 Sep 2010 12:45:34 +0200 > > > Then a few more questions: > > 1) if doubly linked lists really require such pskb_copying, isn't it > > all too costly? > > In the common case the data reference will be one, so we will not > copy. Even if so, one such a case on the fast path should hit performance, so it would need special reviewing. > > > 2) why skb_clone isn't enough instead of pskb_copy? > > Can't share the metadata. I'd really like to understand why the change in handling next/prev should affect more than skb pointers wrt. current sharing. > > > 3) since skb_clone has some cost too, why e.g. saving only the pointer > > to the tail of the list in skb_shared_info isn't enough? > > Then we won't get the rest of the advantages of using list_head such > as prefetching during traversals, automatic debugging facilities, et al. Right, we need to sum pros and cons. So, what's the pros? ;-) Jarek P.