From mboxrd@z Thu Jan 1 00:00:00 1970 From: ebiederm@xmission.com (Eric W. Biederman) Subject: Re: [Patch net-next v2] pktgen: support net namespace Date: Mon, 28 Jan 2013 19:33:58 -0800 Message-ID: <877gmwirx5.fsf@xmission.com> References: <1359424389-21919-1-git-send-email-amwang@redhat.com> <874ni0k965.fsf@xmission.com> <1359428563.20729.12.camel@cr0> Mime-Version: 1.0 Content-Type: text/plain Cc: netdev@vger.kernel.org, "David S. Miller" To: Cong Wang Return-path: Received: from out03.mta.xmission.com ([166.70.13.233]:48580 "EHLO out03.mta.xmission.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752326Ab3A2DeK (ORCPT ); Mon, 28 Jan 2013 22:34:10 -0500 In-Reply-To: <1359428563.20729.12.camel@cr0> (Cong Wang's message of "Tue, 29 Jan 2013 11:02:43 +0800") Sender: netdev-owner@vger.kernel.org List-ID: Cong Wang writes: > On Mon, 2013-01-28 at 18:36 -0800, Eric W. Biederman wrote: >> Cong Wang writes: >> >> > From: Cong Wang >> > >> > v2: remove a useless check >> > >> > This patch add net namespace to pktgen, so that >> > we can use pktgen in different namespaces. >> > >> > Cc: Eric W. Biederman >> > Cc: David S. Miller >> > Signed-off-by: Cong Wang >> > >> > --- >> > net/core/pktgen.c | 123 +++++++++++++++++++++++++++++++++++------------------ >> > 1 files changed, 81 insertions(+), 42 deletions(-) >> >> Skiming through this again I have spotted what looks like a pretty >> major bug. You are limiting yourself to one network device per network >> namespace when the actual limit is one network device per thread. >> >> I think you can just kill the dev member of pktgen_net and the two or >> three lines of code that touch it. > > Good point! > > It is used by pktgen_device_event() to check if the device generates the > event is the one in our namespace. Which of course is trivial with dev_net()...; > It is safe to continue the search even if it is not in our namespace, > but it is not efficient. Probably we need to make pktgen_threads list > per-namespace. Having looked at the code a bit more I think the solution really is to make the proc files per network namespace as you are doing, but to leave the threads per cpu. Then it is just a matter of adding for_each_net loops in the in the paths that add and remove the proc files. The per interface proc files would of course only show up in a single network namespace. It is still a proc file per network namespace and per cpu but that is better than threads that try and consume cpu resources. Especially since the threads when runnign try and consume 100% of the cpu for transmitting packets. Eric