On Wednesday 22 October 2014 18:19:47 Guenter Roeck wrote:
> On Wed, Oct 22, 2014 at 02:29:06PM +0200, Pali Rohár wrote:
> > On Tuesday 21 October 2014 06:27:23 Guenter Roeck wrote:
> > > On 10/20/2014 09:46 AM, Pali Rohár wrote:
> > > > Ok, I will describe my problem. Guenter, maybe you can
> > > > find another solution/fix for it.
> > > > 
> > > > Calling i8k_get_temp(3) on my laptop without
> > > > I8K_TEMPERATURE_BUG always returns value 193 (which is
> > > > above I8K_MAX_TEMP).
> > > > 
> > > > When I8K_TEMPERATURE_BUG is enabled (by default) then
> > > > i8k_get_temp(3) returns value from prev[3] and store new
> > > > value I8K_TEMPERATURE_BUG to prev[3]. Value in prev[3]
> > > > is initialized to 0.
> > > > 
> > > > What I want to achieve is: when i8k_get_temp() for
> > > > particular sensor id always returns invalid value (>
> > > > I8K_MAX_TEMP) then we should totally ignore sensor with
> > > > that id and do not export it via hwmon.
> > > > 
> > > > My solution is: initialize prev[id] to I8K_MAX_TEMP, so
> > > > on invalid data first call to i8k_get_temp(id) returns
> > > > I8K_MAX_TEMP. Then in i8k_init_hwmon check if value is
> > > > < I8K_MAX_TEMP and if not ignore sensor id.
> > > > 
> > > > Guenter, it is clear now? Are you ok that we should
> > > > ignore sensor if always report value above
> > > > I8K_MAX_TEMP? If you do not like my solution/patch for
> > > > it, can you specify how other can it be fixed?
> > > 
> > > I still don't see the point in initializing prev[].
> > 
> > Now prev[] is initialized to 0. It means that first call
> > i8k_get_temp() (with sensor id which return value >
> > I8K_MAX_TEMP) returns 0. Second and other calls returns
> > I8K_MAX_TEMP.
> > 
> > So point is to return same value for first and other calls.
> 
> Yes, I realized that after I sent my previous mail.
> 
> > > Yes, I am ok with ignoring sensor values if the reported
> > > temperature is above I8K_MAX_TEMP. I am just not sure if
> > > we should check against I8K_MAX_TEMP or against, say,
> > > 192. Reason is that we do know that the sensor can
> > > erroneously return 0x99 on some systems once in a while.
> > > We would not want to ignore those sensors just because
> > > they happen to report 0x99 during initialization.
> > > 
> > > So maybe make it
> > > 
> > > 	if (err >= 0 && err < 192)
> > > 
> > > and add a note before the first if(), explaining that
> > > higher values suggest that there is no sensor attached.
> > > 
> > > Thanks,
> > > Guenter
> > 
> > Right, now we need to decide which magic constant to use...
> > 
> > And now I found another problem :-)
> > 
> > On my laptop i8k_get_temp(3) not always return value 193. It
> > is only when AMD graphics card is turned off. When card is
> > on i8k_get_temp(3) returns same value as temperature hwmon
> > part from radeon DRM driver.
> 
> Can you turn the GPU on or off during runtime ?
> That would make it really tricky to handle the situation.
> 

Yes. New laptops with Nvidia Optimus or AMD PowerXpress or Enduro 
technology are designed to automatically turn off secondary GPU 
when is not in use. And nouveau/radeon drivers together with 
vga_switcheroo implements this dynamic power on/off.

> > So it looks like that on my laptop i8k sensor with id 3
> > reports GPU temperature.
> > 
> > When card is turned off radeon driver reports -EINVAL for
> > temperature hwmon sysnode.
> > 
> > So now I think i8k could not ignore sensor totally as it can
> > be mapped to some HW which can be dynamically turned on/off
> > (like my graphics card).
> > 
> > So what do you think about reporting -EINVAL instead
> > I8K_MAX_TEMP when dell SMM returns value above
> > I8K_MAX_TEMP?
> 
> -EINVAL is supposed to mean "Invalid Argument", so it really
> has ia different semantics. We could use -ENXIO, "No such
> device or address", which seems more appropriate.
> 

I prefer to use -EINVAL because other driver (radeon) is using it 
and userspace "sensors" programs handle EINVAL and show "N/A" in 
output instead reporting some error about reading value. If 
nothing else consistency (with other drivers) is my argument.

> Overall, I think the entire error handling is broken and
> should be replaced. One option would be to explicitly check
> for 0x99 and, if detected, go to sleep for, say, 100ms and
> try again. If it still fails, and for all other bad values,
> return -ENXIO. Then the calling code can either return the
> error to user space in the show function, or not install the
> sensor in the probe function.
> 
> Does that make sense ?
> 

Yes, replacing error handling with retry call (after some sleep) 
is better then current code (which returns previous value or 
returns I8K_MAX_TEMP).

But your solution not install the sensor if probe fails on bad 
value does not solve problem that i8k.ko is loading at time when 
GPU card is turned off.

I think current check for installing sensor (err < 0) is enough 
and when invalid value (> I8K_MAX_TEMP) is returned just do not 
show this value for userspace in hwmon sysnode.

> Thanks,
> Guenter

-- 
Pali Rohár
pali.rohar@gmail.com