From mboxrd@z Thu Jan 1 00:00:00 1970 From: Vivien Didelot Subject: Re: [PATCH 1/2] net: dsa: mv88e6xxx: sleep in _mv88e6xxx_stats_wait Date: Fri, 10 Jul 2015 15:21:47 -0400 (EDT) Message-ID: <645518234.216003.1436556107630.JavaMail.zimbra@savoirfairelinux.com> References: <1436547449-26927-1-git-send-email-vivien.didelot@savoirfairelinux.com> <20150710171027.GB6585@groeck-UX31A> <1055594065.210272.1436552447158.JavaMail.zimbra@savoirfairelinux.com> <20150710183623.GB19854@roeck-us.net> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Cc: netdev , Florian Fainelli , David , Andrew Lunn , linux-kernel , kernel To: Guenter Roeck Return-path: Received: from mail.savoirfairelinux.com ([209.172.62.77]:62089 "EHLO mail.savoirfairelinux.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932290AbbGJTVv (ORCPT ); Fri, 10 Jul 2015 15:21:51 -0400 In-Reply-To: <20150710183623.GB19854@roeck-us.net> Sender: netdev-owner@vger.kernel.org List-ID: Hi Guenter, On Jul 10, 2015, at 2:36 PM, Guenter Roeck linux@roeck-us.net wrote: > Hi Vivien, > > On Fri, Jul 10, 2015 at 02:20:47PM -0400, Vivien Didelot wrote: >> > >> > is this really beneficial and/or needed ? >> >> Except using existing generic code, no. >> >> > It adds at least 1ms delay to a loop which did not have any delay at >> > all unless the register read itself was sleeping. >> >> I must have missed where is the benefit from spin reading 10 times this >> register, rather than sleeping 1ms between tests. Does this busy bit >> behaves differently from the phy, atu, scratch, or vtu busy bits? >> > Benefit is reaction time, mostly. If the result isn't ready after the > first spin, the new code path adds a mandatory 1-2ms delay. This could > add up to a lot if that kind of retry is seen a lot. To me, it looks like if this mandatory 1-2ms delay is an issue, then _mv88e6xxx_wait must be fixed. Maybe reducing this delay is an option? > I don't now if there is a specific time limit for this busy bit, > and/or if it behaves differently than the others in terms of timing. > >> > Is the original function seen to return a timeout error under some >> > circumstances ? >> >> I didn't experience it myself, but I guess it may happen. In addition to >> that, the current implementation doesn't check eventual read error. >> That's why I saw a benefit in using _mv88e6xxx_wait(). > > Checking for a read error (or a timeout) is definitely a good thing. > I could also imagine that, for example, a "clear statistics" request > takes more time than currently supported. This is why I asked if you > had seen a timeout with the old code. > > Personally I'd rather leave the wait loop alone and only introduce > error checking unless there is a reason to introduce a sleep, > but I'd like to hear Andrew's and/or Florian's opinion. Andrew may not reply since he's on vacation, but I add Florian in Cc. Thanks, -v