From mboxrd@z Thu Jan 1 00:00:00 1970 From: James Bottomley Subject: Re: [PATCH] scsi: don't count non-failure CHECK_CONDITION as error Date: Fri, 15 Jan 2016 11:36:08 -0800 Message-ID: <1452886568.2356.23.camel@HansenPartnership.com> References: <20160114214602.GC3520@mtj.duckdns.org> <1452872774.2356.5.camel@HansenPartnership.com> <1452873302.2356.9.camel@HansenPartnership.com> <56992361.7030606@interlog.com> <1452882934.2356.13.camel@HansenPartnership.com> <20160115184252.GJ3520@mtj.duckdns.org> <1452884956.2356.21.camel@HansenPartnership.com> <20160115192755.GK3520@mtj.duckdns.org> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Return-path: Received: from bedivere.hansenpartnership.com ([66.63.167.143]:57952 "EHLO bedivere.hansenpartnership.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754515AbcAOTgL (ORCPT ); Fri, 15 Jan 2016 14:36:11 -0500 In-Reply-To: <20160115192755.GK3520@mtj.duckdns.org> Sender: linux-scsi-owner@vger.kernel.org List-Id: linux-scsi@vger.kernel.org To: Tejun Heo Cc: dgilbert@interlog.com, "Martin K. Petersen" , linux-scsi@vger.kernel.org, Dave Jones , kernel-team@fb.com On Fri, 2016-01-15 at 14:27 -0500, Tejun Heo wrote: > Hello, James. > > On Fri, Jan 15, 2016 at 11:09:16AM -0800, James Bottomley wrote: > > If we want to change what is being counted, we have to change the > > definition, so what is the definition you want to see for counting > > errors? and what's the reason driving this change? > > There isn't anything else to it than what I already wrote. It goes > up > for things which aren't failures making the counter essentially > useless. It's confusing from userland as it's unintuitive that these > commands are wrapped in SCSI ATA passthrough commands which use sense > data for result reporting and doing equivalent operations on native > SCSI and ATA devices lead to different outcomes. The counter as-is > is > just useless for libata devices. If it can be rectified easily, > great. If not, it isn't anything critical. That's the point we were making: it can't be rectified easily. Simply counting non-good returns is easy. Counting something that's closer to errors is hard because first we have to define what an "error" is (I could see queue full being not an error, but what about not ready or initializing command required?) and then implement and infrastructure to classify and count the returns, which means an awful lot of sense data parsing. James