From mboxrd@z Thu Jan 1 00:00:00 1970 From: linas@austin.ibm.com (Linas Vepstas) Subject: Re: [PATCH]: e1000: prevent statistics from getting garbled during reset. Date: Fri, 31 Mar 2006 11:03:19 -0600 Message-ID: <20060331170319.GV2172@austin.ibm.com> References: <20060330213928.GQ2172@austin.ibm.com> <20060331000208.GS2172@austin.ibm.com> <442C8069.507@wolfmountaingroup.com> <20060331003506.GU2172@austin.ibm.com> <442CACC0.1060308@wolfmountaingroup.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: john.ronciak@intel.com, jesse.brandeburg@intel.com, jeffrey.t.kirsher@intel.com, Jeff Garzik , linux-kernel@vger.kernel.org, netdev@vger.kernel.org, linux-pci@atrey.karlin.mff.cuni.cz, linuxppc-dev@ozlabs.org Return-path: To: "Jeffrey V. Merkey" Content-Disposition: inline In-Reply-To: <442CACC0.1060308@wolfmountaingroup.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Thu, Mar 30, 2006 at 09:14:56PM -0700, Jeffrey V. Merkey wrote: > Yes, we need one. The adapter needs to maintain these stats from the > registers in the kernel structure and not > its own local variables. Did you read the code to see what the adapter does with these stats? Among other things, it uses them to adaptively modulate transmit rates to avoid collisions. Just clearing the hardware-private stats will mess up that function. > That way, when someone calls to clear the stats > for testing and analysis purposes, > they zero out and are reset. 1) ifdown/ifup is guarenteed to to clear things. Try that. 2) What's wrong with taking deltas? Typical through-put performance measurement is done by pre-loading the pipes (i.e. running for a few minutes wihtout measuring, then starting the measurement). I'd think that snapshotting the numbers would be easier, and is trivially doable in user-space. I guess I don't understand why you need a new kernel featre to imlement this. --linas