From mboxrd@z Thu Jan 1 00:00:00 1970 From: Satoshi OSHIMA Subject: Re: [RFC/PATCH 3/3] UDP memory usage accounting (take 2): measurement Date: Mon, 01 Oct 2007 22:52:27 +0900 Message-ID: <4700FB9B.1060304@hitachi.com> References: <46FD048B.9010409@hitachi.com> <20070928143710.GA16747@2ka.mipt.ru> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Cc: Andi Kleen , netdev , =?windows-1252?Q?=3F=3F_=3F=3F?= , Yumiko SUGITA , =?windows-1252?Q?=22=3F=3F=40RedHat=22?= , David Miller , Herbert Xu To: Evgeniy Polyakov Return-path: Received: from mail7.hitachi.co.jp ([133.145.228.42]:39341 "EHLO mail7.hitachi.co.jp" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752161AbXJANwd (ORCPT ); Mon, 1 Oct 2007 09:52:33 -0400 Received: from mlsv16.hitachi.co.jp (unknown [133.144.234.166]) by mail7.hitachi.co.jp (Postfix) with ESMTP id C2A5B37ADE for ; Mon, 1 Oct 2007 22:52:31 +0900 (JST) In-Reply-To: <20070928143710.GA16747@2ka.mipt.ru> Sender: netdev-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Evgeniy Polyakov wrote: > On Fri, Sep 28, 2007 at 10:41:31PM +0900, Satoshi OSHIMA (satoshi.oshima.fk@hitachi.com) wrote: >> This patch introduces memory usage measurement for UDP. >> >> These 3 points were updated. >> >> - UDP specific codes in IP layer were removed. >> >> - atomic_sub() in a loop was removed >> >> - accounting during socket destruction > > Another approach is to account only at the highest UDP layer and having > datagram skb destructor just like it is done in TCP, but this approach > is also resonable. This patch set try to introduce a memory accounting by the page because TCP does. And ip_append_data() merges payloads to a sk_buff if previous sk_buff has enough space. The problem is that udp_append_data() doesn't recognize whether this merge happens or not. If the accounting must be in UDP layer, we need to change the interface of ip_append_data() to know this merge happens. Once the interface is changed, we have to maintain other protocol stacks to keep up with the change. But I didn't want to do it to keep this patch set small in the first step. > I already told that patches 1 and 3 have broken indent, please fix that. Oops! I will fix that. > A hint: when you are about to submit something network related for inclusion, > and strongly believes it is ready, it can be a not that bad idea to add > David Miller to copy list, he can complain about > backlog and so on, but will read you mail twice :) but do not tell anyone. Thank you for your advice. I will do that! Satoshi Oshima