From mboxrd@z Thu Jan 1 00:00:00 1970 From: Lennert Buytenhek Subject: Re: [PATCH] use mmiowb in tg3_poll Date: Sun, 29 May 2005 01:12:10 +0200 Message-ID: <20050528231209.GA8456@xi.wantstofly.org> References: <200410211628.06906.jbarnes@engr.sgi.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: netdev@oss.sgi.com, jbarnes@engr.sgi.com, gnb@sgi.com Return-path: To: akepner@sgi.com Content-Disposition: inline In-Reply-To: Sender: netdev-bounce@oss.sgi.com Errors-to: netdev-bounce@oss.sgi.com List-Id: netdev.vger.kernel.org On Fri, Oct 22, 2004 at 01:51:01PM -0700, akepner@sgi.com wrote: > Returning from tg3_poll() without flushing the PIO write which > reenables interrupts can result in lower cpu utilization and higher > throughput. I'm quite curious what kind of MMIO read latency you see on your Altix boxen. This app is quite useful for determining those figures on x86{,_64} machines: http://svn.gnumonks.org/trunk/mmio_test/mmio_test.c I'm not sure if it'll run on Itanium out of the box (it currently assumes PAGE_SIZE <= 4096, you'd probably need to tweak rdtscl(), and if there's a weird phys:pci address correspondence you might have to teach it about that as well), but it would be nice if you could give an approximate indication of exactly how expensive an MMIO read is on your platform. cheers, Lennert