From mboxrd@z Thu Jan 1 00:00:00 1970 From: Po-Yu Chuang Subject: Re: [PATCH v4] net: add Faraday FTMAC100 10/100 Ethernet driver Date: Thu, 24 Feb 2011 15:27:55 +0800 Message-ID: References: <20110131.203556.193730771.davem@davemloft.net> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Cc: mirqus@gmail.com, netdev@vger.kernel.org, linux-kernel@vger.kernel.org, bhutchings@solarflare.com, eric.dumazet@gmail.com, joe@perches.com, dilinger@queued.net, ratbert@faraday-tech.com To: David Miller Return-path: In-Reply-To: <20110131.203556.193730771.davem@davemloft.net> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Hi David, On Tue, Feb 1, 2011 at 12:35 PM, David Miller wrote: > From: Po-Yu Chuang > Date: Tue, 1 Feb 2011 11:56:16 +0800 > >> If I simply allocate a page for each rx ring entry, I still need to allocate >> an skb and copy at least packet header in first page to skb->data. Then >> add the page of rest of payload to skb by skb_fill_page_desc(). > > You should attach the pages, the use __pskb_pull_tail() to bring in the > headers to the linear skb->data area. > > See drivers/net/niu.c:niu_process_rx_pkt(). I tried two ways to implement zero-copy. One is to preallocate skb big enough for any rx packet and use the skb as rx buffer. The other is use page as rx buffer, use skb_fill_page_desc() to add a data page to skb and then pull only header to skb by __pskb_pull_tail() as you suggested. Two implementations are slower than the original memcpy version. (benchmarked with iperf) I guess the problem is because a HW restriction that the rx buffer must be 64 bits aligned. Since I cannot make rx buffer starts at offset 2 bytes, the IP header, TCP header and data are not 4 bytes aligned. The performance drops drastically. Therefore, I will submit later a v6 which is still using memcpy(). best regards, Po-Yu Chuang