From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jason Wang Subject: Re: [RFC PATCH] virtio_net: XDP support for adjust_head Date: Wed, 4 Jan 2017 11:22:46 +0800 Message-ID: <40812dd8-3fa2-ed85-9c99-d4b22a3d7b31@redhat.com> References: <20170102194413.9089.39078.stgit@john-Precision-Tower-5810> <73715f7a-eeeb-679f-a7b8-7b1fefe1757e@redhat.com> <586BD734.7020105@gmail.com> <586BD7F3.60109@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 8bit Cc: john.r.fastabend@intel.com, netdev@vger.kernel.org, alexei.starovoitov@gmail.com, daniel@iogearbox.net To: John Fastabend , mst@redhat.com Return-path: Received: from mx1.redhat.com ([209.132.183.28]:44850 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1761303AbdADDaC (ORCPT ); Tue, 3 Jan 2017 22:30:02 -0500 In-Reply-To: <586BD7F3.60109@gmail.com> Sender: netdev-owner@vger.kernel.org List-ID: On 2017年01月04日 00:57, John Fastabend wrote: >>>> + /* Changing the headroom in buffers is a disruptive operation because >>>> + * existing buffers must be flushed and reallocated. This will happen >>>> + * when a xdp program is initially added or xdp is disabled by removing >>>> + * the xdp program. >>>> + */ >>> We probably need reset the device here, but maybe Michale has more ideas. And if >>> we do this, another interesting thing to do is to disable EWMA and always use a >>> single page for each packet, this could almost eliminate linearizing. >> Well with normal MTU 1500 size we should not hit the linearizing case right? The >> question is should we cap the MTU at GOOD_PACKET_LEN vs the current cap of >> (PAGE_SIZE - overhead). > Sorry responding to my own post with a bit more detail. I don't really like > going to a page for each packet because we end up with double the pages in use > for the "normal" 1500 MTU case. We could make the xdp allocation scheme smarter > and allocate a page per packet when MTU is greater than 2k instead of using the > EWMA but I would push those types of things at net-next and live with the > linearizing behavior for now or capping the MTU. > Yes, agree. Thanks