From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jean Delvare Subject: Re: [PATCH] e1000: Real time packets and bytes statistics Date: Thu, 12 Oct 2006 23:02:52 +0200 Message-ID: <20061012230252.8b5a1ebb.khali@linux-fr.org> References: <20061011133557.8a404e24.khali@linux-fr.org> <4807377b0610111044u5eced127od8272797e63743a7@mail.gmail.com> <20061012103006.1367693a@freekitty> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: "Jesse Brandeburg" , netdev@vger.kernel.org Return-path: Received: from smtp-104-thursday.noc.nerim.net ([62.4.17.104]:34309 "EHLO mallaury.nerim.net") by vger.kernel.org with ESMTP id S1750804AbWJLVCs (ORCPT ); Thu, 12 Oct 2006 17:02:48 -0400 To: Stephen Hemminger In-Reply-To: <20061012103006.1367693a@freekitty> Sender: netdev-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Hi Stephen, On 10/11/06, Stephen Hemminger wrote: > On Wed, 11 Oct 2006, Jesse Brandeburg wrote: > > On 10/11/06, Jean Delvare wrote: > > > Let the e1000 driver report the most important statistics (rx/tx_bytes > > > and rx/tx_packets) in real time, rather than every other second. This > > > is similar to what the e100 driver is doing. > > > (...) > > > --- linux-2.6.19-rc1.orig/drivers/net/e1000/e1000_main.c 2006-10-11 10:53:49.000000000 +0200 > > > +++ linux-2.6.19-rc1/drivers/net/e1000/e1000_main.c 2006-10-11 11:34:41.000000000 +0200 > > > @@ -3118,6 +3118,8 @@ > > > e1000_tx_map(adapter, tx_ring, skb, first, > > > max_per_txd, nr_frags, mss)); > > > > > > + adapter->net_stats.tx_packets++; > > > + adapter->net_stats.tx_bytes += skb->len; > > > netdev->trans_start = jiffies; > > > > this is the part I'm most worried about. as I believe it to be > > incorrect for TSO packets. Maybe something like? > > + if (skb_shinfo(skb)->gso_segs) > > + adapter->net_stats.tx_packets += skb_shinfo(skb)->gso_segs; > > + else > > + adapter->net_stats.tx_packets++; > > + adapter->net_stats.tx_bytes += skb->len; > > netdev->trans_start = jiffies; > > > > skb len will still be off by some amount, because the skb->data > > (header) is replicated across each gso segment but only counted once > > this way, but hopefully someone will pipe up with a good way to > > compute that. > > You might want to put the tx values in a per-cpu structure and > sum later. Incrementing statistics can actually be a performance > bottleneck on SMP tests, because it causes lots of cache thrashing. I don't really see how this would be implemented. Can you please point me to other drivers which do it that way? Thanks, -- Jean Delvare