On Fri, Jan 31, 2014 at 09:17:29AM -0800, Sören Brinkmann wrote: > Hi Guenter, > > On Thu, Jan 30, 2014 at 08:01:02PM -0800, Guenter Roeck wrote: > > Hi Sören, > > > > On Thu, Jan 30, 2014 at 05:13:42PM -0800, Sören Brinkmann wrote: > > > Hi Guenther, > > > > > > I recently became aware of some weird behavior when using the UCD9000 > > > driver for voltage monitoring. We use that device on our zc706 Zynq > > > platform to generate 5 voltage rails. Observing the voltages in Linux > > > showed a correct value for rail 1 but a 2x error in in rails 2-3 and a > > > 4x error in rail 5. Digging through specs, docs and sources I could > > > track this down to this: > > > - The PMBus core assumes one VOUT_MODE register per chip > > > - UCD90120 has VOUT_MODE registers per rail > > > > > > In our setup we have an exponent of -14 for rail 1, -13 for rails 2-4 and > > > -14 for rail 5. The PMBus core only reads and uses the first one and > > > applies it to all rails. > > > > > > I'm not sure whether the UCD devices don't comply to spec or whether > > > this is a bug in the PMBus core, either way, the reported values from > > > the driver are prone to be wrong/scaled by an unknown factor. > > > > > The PMBus specification is sufficiently vague to ensure that pretty much > > any interpretation of it is PMBus compliant. > > > > > Is this somehow known behavior for which a workaround might be > > > available? Otherwise, if you know how an acceptable solution might look > > > like, I might find some time and give it a shot. > > > > > I think we may have to expand the driver to detect and configure the exponent > > per rail/page. At least I can not think of another solution right now. > > > > Can you test the patch below ? Completely untested, but it compiles. > > It will likely need some refinement (for example, other chips may not even > > have the VOUT_MODE command on non-zero pages), but should be useful for > > testing. > > That was fast, thanks. The patch works. With it, the driver reports the > correct voltages. > > I suspect that the UCD9200 were affected as well. Unfortunately, in our > configs for those the exponents across all rails are the same, so I > don't see things go wrong. > I have eval boards for several ucd92xx chips, so I should be able to test that. I was not aware that the exponent is configurable and depends on the rail, so I completely neglected that. > You can add my > Tested-by: Sören Brinkmann <soren.brinkmann@xxxxxxxxxx> > Thanks a lot for testing! > Let me know if you need a different patch to be tested. > I'll have to make sure that the new code doesn't break the driver for any other chips using linear mode. I hope I can get that done over the weekend. I'll let you know how it goes. Thanks, Guenter _______________________________________________ lm-sensors mailing list lm-sensors@xxxxxxxxxxxxxx http://lists.lm-sensors.org/mailman/listinfo/lm-sensors
