From: arnd@arndb.de (Arnd Bergmann) Date: Fri, 8 Apr 2011 17:02:17 +0200 Subject: [RFC PATCHv2 1/2] Export SoC info through sysfs In-Reply-To: <4D9EBC2C.2050809@linaro.org> References: <1299846911-15782-1-git-send-email-maxime.coquelin-nonst@stericsson.com> <201104080535.02769.arnd@arndb.de> <4D9EBC2C.2050809@linaro.org> Message-ID: <201104081702.17409.arnd@arndb.de> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On Friday 08 April 2011, Lee Jones wrote: > On 08/04/11 04:35, Arnd Bergmann wrote: > > On Friday 08 April 2011, Greg KH wrote: > >> Symlinks are a requirement as multiple cpus can be attached to a single > >> SoC. > >> > >> What about multiple cpus that are attached to multiple SoCs? Why even > >> try to describe this relationship, what would userspace get out of this > >> information? > > > > The only one I can think of is node affinity. I've worked with a system > > (IBM QS2x Cell blade) that had two SoCs with multiple CPUs each. There was > > a significant performance penalty when talking to devices on the remote > > SoC. In that case, we used the NUMA node information in sys to deal > > with it, but we might not want to use the NUMA infrastructure on systems > > that only have RAM on one node. > > I'm struggling to see a scenario where we'd have multiple SoCs on a > single device and only one filesystem. ux500 sub-arch code is yet to > facilitate SoC counting functionality. If were are to implement per-SoC > directories would this need to be added? How can you say we won't need it when I've just given you a specific example of a widely shipping machine? It's not important that the kernel internal code supports multiple SoCs as long as there is no in-tree user for this. It can always be added at a later point. What is important is that the user interface accomodates for it. If you name the device simply /sys/devices/soc/, any user program looking at that will have to be modified when we have more than one. We cannot allow that to happen. Arnd