Return-Path: Subject: Re: [PATCH 1/2] i2c: mux: pca954x: change the default to deselect after each transfer References: <20190121164742.35718-1-robertshearman@gmail.com> <9b484cde-c20c-2078-9a55-eab27b56ef36@axentia.se> From: Robert Shearman Message-ID: <66870ca6-4789-57d9-ce30-2beb96efb86b@gmail.com> Date: Tue, 22 Jan 2019 10:17:43 +0000 MIME-Version: 1.0 In-Reply-To: <9b484cde-c20c-2078-9a55-eab27b56ef36@axentia.se> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-GB Content-Transfer-Encoding: 7bit To: Peter Rosin , Rob Herring , Mark Rutland Cc: "linux-i2c@vger.kernel.org" , "devicetree@vger.kernel.org" , Robert Shearman List-ID: Hi Peter, On 22/01/2019 08:41, Peter Rosin wrote: > Hi Robert, > > Thanks, for your patches. > > On 2019-01-21 17:47, Robert Shearman wrote: >> From: Robert Shearman >> >> The behaviour, by default, to not deselect after each transfer is >> not safe/correct when there is a device with an address that conflicts >> with another device either on the parent bus, or on another pca954x >> mux on the same parent bus. > > Right. The current way might not be the safest, but the way I see it, > this ship has sailed a long time ago. It might take quite a while > before someone reports a regression, but when that happens and we > need to revisit or even revert this, things will get ugly. > > Also, doubling or even tripling the I2C traffic (most I2C xfers > are small) for unsuspecting users with more normal I2C setups > (w/o multiple muxes) just to accommodate the more complexes cases > is arguably not the correct balance. Ok, thanks for considering the changes. Would a device-level sysfs toggle or module parameter approach be more acceptable? The use case where we are hitting this issue is on an x86_64 ACPI platform where we desire to use a general purpose kernel, so the existing devicetree and platform data toggle approaches are not ideal for us. > >> Therefore, default to the least surprising mode, where the mux >> channels are deselected after each transaction, which is safe in the >> face of one or more devices hanging off the mux with addresses that >> conflict with other devices on different muxes, or on the parent > > Side note, and not that it matters since it's a NACK anyway, but > having an address conflict with the parent bus is not allowed and is > just bad hardware design that simply cannot work reliably, if at all. Yes, good point. Thanks, Rob