Ticket 1552

Mark Studebaker mds4 at verizon.net
Tue Feb 3 03:30:13 CET 2004


Back in late 2002/Jan. '03 we had a lot of discussion about read errors.
Certainly the situation then and now, where most drivers just return a -1,
isn't great.
One sentiment was that drivers definitely shouldn't silently return
the old value, that definitely wasn't "server-class".
People felt strongly about this.

So what I did back then was update the adm1021 driver as a demonstration.
This was just after it got ported to 2.6, so my changes are only
in our package, not 2.6.
This driver now sets the appropriate alarms bit on a read failure,
and returns the old value. Take a look.

I think at the time I also proposed a new /proc entry, similar to
alarms and with the same bitmask, called 'fail'. This would
provide an indication separate from the alarm indication.
I'm sure there are other ways... perhaps sysfs opens some new possibilities.
Any way to make the sysfs read fail, for example?

You may wish to check the mail archives for more info.


mds

Jean Delvare wrote:
>>Ok, I've patched the module again and adapted the script a bit -
>>adding timestamps & debug log retrieval.
> 
> 
> You're quite good at Perl, aren't you? :)
> 
> 
>>I have uploaded the first output with 2 seconds between execution to:
>>      http://james.evilpenguin.com/sensors/
>>
>>I thought 1000 executions would be sufficient unless you think I need
>>to run it longer or for a specified period of time?
> 
> 
> No, it's perfect.
> 
> Strange that the number of errors varies so much between the three runs,
> with no obvious logic.
> 
> Here are the stats:
> Immediate success: 85.5%
> Success after 1 retry: 10.4%
> Success after 2 retries: 0.9%
> Success after 3 retries: 0.03%
> 
> The high ratio of non-immediate successes clearly shows that the new
> read method is welcome. The fact 3 retries were necessary only once out
> of 3000 reads shows that setting max_retries to 5 isn't a bad choice. I
> think we can leave it this way.
> 
> I'm pretty happy with the results and think of sending the patch to Greg
> KH for inclusion into Linux 2.6, if it's OK for you too. I also plan to
> backport the changes to our CVS (Linux 2.4 compatible) driver.
> 
> Thanks a lot for the great testing of my code. You'll got credits.
> 



More information about the lm-sensors mailing list