From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alexei Starovoitov Subject: Re: [RFC PATCH 2/2] ixgbe: add af_packet direct copy support Date: Mon, 30 Jan 2017 18:53:05 -0800 Message-ID: <588FFC11.9030403@fb.com> References: <20170127213132.14162.82951.stgit@john-Precision-Tower-5810> <20170127213411.14162.22553.stgit@john-Precision-Tower-5810> Mime-Version: 1.0 Content-Type: text/plain; charset="utf-8"; format=flowed Content-Transfer-Encoding: 7bit Cc: , To: John Fastabend , , , , Return-path: Received: from mx0a-00082601.pphosted.com ([67.231.145.42]:37817 "EHLO mx0a-00082601.pphosted.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750929AbdAaEMU (ORCPT ); Mon, 30 Jan 2017 23:12:20 -0500 In-Reply-To: <20170127213411.14162.22553.stgit@john-Precision-Tower-5810> Sender: netdev-owner@vger.kernel.org List-ID: On 1/27/17 1:34 PM, John Fastabend wrote: > + h2 = page_address(rx_buffer->page) + rx_buffer->page_offset - hdrlen; > + eth = page_address(rx_buffer->page) + rx_buffer->page_offset, I don't think it compiles ;) > + /* This indicates a bug in ixgbe leaving for testing purposes */ > + WARN_ON(TP_STATUS_USER & h2->tp_status); > + len = le16_to_cpu(rx_desc->wb.upper.length); > + h2->tp_len = len; > + h2->tp_snaplen = len; > + h2->tp_mac = ALIGN(TPACKET_ALIGN(TPACKET2_HDRLEN), L1_CACHE_BYTES); > + h2->tp_net = h2->tp_mac + ETH_HLEN; > + h2->tp_sec = div_s64_rem(ns, NSEC_PER_SEC, &rem); > + h2->tp_nsec = rem; > + > + sll = (void *)h2 + TPACKET_ALIGN(sizeof(struct tpacket2_hdr)); > + sll->sll_halen = ETH_HLEN; > + memcpy(sll->sll_addr, eth->h_source, ETH_ALEN); > + sll->sll_family = AF_PACKET; > + sll->sll_hatype = rx_ring->netdev->type; > + sll->sll_protocol = eth->h_proto; > + sll->sll_pkttype = PACKET_HOST; > + sll->sll_ifindex = rx_ring->netdev->ifindex; performance wise it looks very expensive to do all these header copies and integer divide for every packet. I think unless we move to new dumb and simple header format performance of this approach is not going to be satisfactory.