From owner-freebsd-net Wed Jan 2 8:10:44 2002 Delivered-To: freebsd-net@freebsd.org Received: from smtp.www-service.de (smtp.www-service.de [212.77.161.16]) by hub.freebsd.org (Postfix) with ESMTP id A327A37B42F for ; Wed, 2 Jan 2002 08:10:25 -0800 (PST) Received: from fw.tue.le (pD95033EC.dip.t-dialin.net [217.80.51.236]) by smtp.www-service.de (8.11.6/8.11.6) with ESMTP id g02HWOI28459; Wed, 2 Jan 2002 18:32:25 +0100 Received: from mezcal.tue.le (mezcal.tue.le [192.168.201.20]) by fw.tue.le (8.8.8/8.8.8) with ESMTP id RAA08017; Wed, 2 Jan 2002 17:09:17 +0100 (CET) (envelope-from thz@mezcal.tue.le) Received: (from thz@localhost) by mezcal.tue.le (8.11.6/8.11.6) id g02G9Hj08358; Wed, 2 Jan 2002 17:09:17 +0100 (CET) (envelope-from thz) Date: Wed, 2 Jan 2002 17:09:17 +0100 From: Thomas Zenker To: Mike Silbersack Cc: net@freebsd.org, Josef Karthauser , Matthew Dillon Subject: Re: USB ethernet problem Message-ID: <20020102170917.A8308@mezcal.tue.le> References: <20011221115851.A10172@mezcal.tue.le> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline User-Agent: Mutt/1.2.5i In-Reply-To: ; from silby@silby.com on Fri, Dec 21, 2001 at 01:35:36PM -0500 Sender: owner-freebsd-net@FreeBSD.ORG Precedence: bulk List-ID: List-Archive: (Web Archive) List-Help: (List Instructions) List-Subscribe: List-Unsubscribe: X-Loop: FreeBSD.org On Fri, Dec 21, 2001 at 01:35:36PM -0500, Mike Silbersack wrote: > On Fri, 21 Dec 2001, Thomas Zenker wrote: > > Might there be a run condition, interrupt problem or priority problem > > with the ata driver, which become visible now on the slower machine? > > > > May be I have to go back to fbsd 4.3 > > Hmm. I agree, it doesn't like like much of anything usb-related changed > between 4.3 and now. If we examine the ata driver, one big thing changed > between 4.3 and 4.4: Write caching was re-enabled. This caused a big > change in disk throughput for some people, perhaps it's affecting you. > Try turning it off (see the ata manpage for details) and see if that > changes your results. > > Also, is your USB adapter sharing interrupts with anything else? Are > interrupts being allocated the same way on 4.3 and 4.4? > > Mike "Silby" Silbersack On Tue, Jan 01, 2002 at 08:32:09PM -0500, Mike Silbersack wrote: > Wow, talk about Terry getting off-topic again... > > Anyway, out of curiosity, did anyone figure out anything about what is > wrong with USB ethernet? I have this feeling that it's just a driver bug, > but since so few developers use usb ethernet cards, it hasn't been > noticed. > > Thomas, you noted that your USB ethernet performance is worse with 4.4 > than it was with 4.3. Could you setup a test where you use a 4.3/usb > client and then 4.3 & 4.4 pci servers? (With the server being the one > primarily sending.) This would be useful in telling us if the changes > which affect you are tcp stack changes, or if something in the usb > subsystem changed. Equalize the slowstart flightsizes and receive / send > buffers in all test cases, of course. > > Thanks, > > Mike "Silby" Silbersack Hi Mike, back from holidays... because this is now discussed in different threads, on -stable and on -net, I will try to recapitulate what has happened and keep this on -net "USB ethernet problem". The performance problems apeared after updating my test system from 4.3 to 4.4 with Netgear EA101 USB/ethernet adaptor (kue driver). Performance dropped by a factor of 10 or more. The server in all cases was 4.4. After testing different slowstart flightsizes and send/recv buffer sizes with ttcp the findings were, that mostly the recvspace reduced to 16K (as in 4.3) recovered the performance. See also my message to -stable from 2001/12/14. The usb host controller on this system is a VIA 83C572 (UHCI). Now going to the final embedded hardware, the suprise was a hanging usb driver. The strange thing is, this does not happen while testing with ttcp, but only if the data is written to disk. The following kernel messages are printed: usb0: host controller process error usb0: host controller halted kue0: watchdog timeout kue0: usb error on tx: TIMEOUT this comes from uhci_intr() in dev/usb/uhci.c. Aparently the usb0-messages reflect a hw status register!? This happens very quickly with 4.4 (it is impossible to install over usb/ethernet), but I have seen it today for the first time with 4.3 also. The usb host controller is UHCI Intel 82371AB/EB (PIIX4). On none of the systems is the interrupt shared. Interrupts are allocated the same way on 4.3 and 4.4. hw.ata.wc has no influence. It looks like a longstanding, but not discovered issue. as for the performance tests: the test from 2001/12/14 (-stable) were done with 4.4&4.3 as client, server was 4.4 - primarily sending. Do you mean to repeat the tests with a 4.3 server? -- Thomas Zenker c/o Lennartz electronic GmbH Bismarckstrasse 136, D-72072 Tuebingen, Germany Phone: +49-(0)7071-93550 Email: thz@lennartz-electronic.de To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-net" in the body of the message