From mboxrd@z Thu Jan 1 00:00:00 1970 From: Guenter Roeck Subject: Re: [PATCH] eeprom: at24: Add support for large EEPROMs connected to SMBus adapters Date: Wed, 25 Mar 2015 07:11:31 -0700 Message-ID: <5512C213.7030705@roeck-us.net> References: <20150205002630.GA396@roeck-us.net> <20150205144028.GA4865@katana> <20150205175326.GA26691@roeck-us.net> <20150212040126.GA1691@roeck-us.net> <20150216120951.GA2840@katana> <20150317042049.GA6765@roeck-us.net> <20150318132707.GD3580@katana> <550A4162.8000009@roeck-us.net> <20150319081612.GA900@katana> <20150319174314.GA17329@roeck-us.net> <20150319213937.GA899@katana> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20150319213937.GA899@katana> Sender: linux-kernel-owner@vger.kernel.org To: Wolfram Sang Cc: linux-i2c@vger.kernel.org, linux-kernel@vger.kernel.org List-Id: linux-i2c@vger.kernel.org On 03/19/2015 02:39 PM, Wolfram Sang wrote: > >> Turns out this is really easy to reproduce. One process reads >> the eeprom over and over again, another runs i2cdump in a loop, >> and voila ... lots of corruptions. Scary, especially considering >> how wide-spread this kind of i2c access is in the kernel. > > A coccinelle script should at least be able to find vulnerable code > paths, maybe even fix it. But not today for me... Thanks for testing and > sharing the results! > Wolfram, just to give you an update: I do have some code, but it is a bit messy, and it doesn't work well for ds2482 (the chip behind it still hangs up if I access it in parallel through i2c-dev). On top of that, it causes pretty significant slow-downs when accessing other devices on the same bus at the same time. Not surprising, I guess, since it expands the scope of the bus lock significantly. I thought about introducing a client lock, but that does not work because of the way i2c-dev is written (creating its own 'shadow' client structure). An address lock (ie a client lock based on instead of one residing in the client structure) seems to be too expensive. So right now I don't really know how to proceed, or if to proceed at all. I'll think about it some more, but given how wide-spread the problem is in the kernel, I might just leave it alone, and keep the at24 changes out of tree. Ultimately, the real problem is how i2c-dev accesses a client, not how i2c client drivers (who assume they have exclusive access to a chip) handle multi-command sequences. Forcing extensive locking on all drivers because of i2c-dev just doesn't seem to be the right thing to do. Guenter