Re: [PATCH v2] platform/x86: add Gigabyte WMI temperature driver

[Guenter Roeck <linux@roeck-us.net>]

On 4/8/21 2:36 AM, Hans de Goede wrote:
> Hi,
> 
> On 4/7/21 9:43 PM, Thomas Weißschuh wrote:
>> Hi Hans,
>>
>> On Mi, 2021-04-07T17:54+0200, Hans de Goede wrote:
>>> Thank you for your new driver and thank you for the quick respin
>>> addressing Barnabás' request to make it a WMI driver.
>>>
>>> The code looks good, so merging this should be a no-brainer,
>>> yet I'm not sure if I should merge this driver as-is, let me
>>> explain.
>>
>> thanks for the encouraging words.
>>
>>> The problem is that I assume that this is based on reverse-engineering?
>>
>> Yes, it is completely reverse-engineered.
>> Essentially I stumbled upon Matthews comment at
>> https://bugzilla.kernel.org/show_bug.cgi?id=204807#c37 and went from there.
>>
>>> We have some mixes experiences with reverse-engineered WMI drivers,
>>> sometimes a feature is not supported yet the wmi_evaluate_method()
>>> call happily succeeds. One example of this causing trouble is:
>>>
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=1797d588af15174d4a4e7159dac8c800538e4f8c
>>
>> There actually are reports of recent, similar mainboards with recent firmware and
>> similar sensor chips that do not support the temperature query.
>> (https://github.com/t-8ch/linux-gigabyte-wmi-driver/issues/3 and
>> https://github.com/t-8ch/linux-gigabyte-wmi-driver/issues/2)
>>
>> Unfortunately for unknown WMI queries the firmware does not return any value.
>> This ends up as an ACPI integer with value 0 on the driver side.
>> (Which I could not yet find documentation for if that is expected)
>> In the current version of the driver EIO is returned for 0 values which
>> get translated to N/A by lm-sensors.
>>
>>> At a minimum I think your driver should check in its
>>> probe function that
>>>
>>> gigabyte_wmi_query_integer(GIGABYTE_WMI_TEMPERATURE_QUERY, ...)
>>>
>>> actually succeeds on the machine the driver is running on chances
>>> are that Gigabyte has been using the DEADBEEF-2001-0000-00A0-C90629100000
>>> GUID for ages, where as the 0x125 value for GIGABYTE_WMI_TEMPERATURE_QUERY
>>> suggests that this is a pretty new API.
>>
>> Would it be enough to probe all six sensors and check if all return 0?
> 
> I think that failing the probe with -ENODEV, with a dev_info() explaining why when
> all six sensors return 0 would be good yes, that should also fix those 2
> issues on https://github.com/t-8ch/linux-gigabyte-wmi-driver/.
> 
>>> It would be good if you can see if you can find some DSDT-s for older
>>> gigabyte motherboards attached to various distro's bug-tracking systems
>>> or forum posts and see how those respond to an unknown gigabyte_wmi_commandtype.
>>
>> Will do.
> 
> Since you alreayd have bugreports of boards where this does not work,
> please don't spend too much time on this. I guess those older DSDT-s will
> also just return an integer with value 0.
> 
>>> Another open question to make sure this driver is suitable
>>> as a generic driver (and does not misbehave) is how to figure out
>>> how many temperature sensors there actually are.
>>
>> So far I could not find out how to query this from the firmware.
>> The IT8688 chip can report the state of each sensor but that is not exposed by
>> the firmware.
>> But even the state information from the IT8688 is not accurate as is.
>> One of the sensors that is reported as being active (directly via it87) on my
>> machine always reports -55°C (yes, negative).
> 
> Ok.
> 
>>> Perhaps the WMI interface returns an error when you query an out-of-range
>>> temperature channel?
>>
>> Also "0" as mentioned above.
> 
> Hmm, so maybe this can be used to limit the amount of reported temperature
> sensors, IOW if sensors 5 and 6 report 0, only register 4 sensors ?
> 
>>
>>> One option here might be to add a DMI matching table and only load on
>>> systems on that table for now. That table could then perhaps also provide
>>> labels for each of the temperature channels, which is something which
>>> would be nice to have regardless of my worries about how well this driver
>>> will work on motherboards on which it has not been tested.
>>
>> I am collecting reports for working motherboards at
>> https://github.com/t-8ch/linux-gigabyte-wmi-driver/issues/1 .
> 
> Good, you should probably ask reporters there to provide the output of:
> 
> grep . /sys/class/dmi/id/* 2> /dev/null
> 
> Ran as a normal user (so that the serial-numbers will be skipped) so that
> you will have DMI strings to match on if you decide to go that route.
> 
>>
>>> You could combine this DMI matching table with a "force" module option to
>>> continue with probing on boards which are not on the table to allow users
>>> to test and report their results to you.
>>>
>>> And hopefully after a while, when we're confident that the code works
>>> well on most gigabyte boards we can drop the DMI table, or at least
>>> only use it for the channel labels.
>>
>> That sounds good.
>>
>>> Please don't take this the wrong way; I think it is great that you are
>>> working on this. And the quick turnaround of the v2 of this drivers makes
>>> me pretty certain that we can figure something out and get this merged.
>>
>> Thank you for the feedback!
>>
>>> Have you tried contacting Gigabyte about this? I don't think the WMI
>>> interface is something which they need to keep secret for competitive
>>> reasons, so maybe they can help? Note if they want you to sign a NDA
>>> of sorts to view docs, then make sure that it contains some language
>>> about them allowing you to release an opensource driver for their
>>> hardware based on the "protected" information.
>>
>> I have not contacted them yet, will do.
>>
>> As mentioned in the initial patch submission there would be different ways to
>> access this information firmware:
>>
>> * Directly call the underlying ACPI methods (these are present in all so far
>>   observed firmwares, even if not exposed via WMI).
>> * Directly access the ACPI IndexField representing the it87 chip.
>> * Directly access the it87 registers while holding the relevant locks via ACPI.
>>
>> I assume all of those mechanisms have no place in a proper kernel driver but
>> would like to get your opinion on it.
> 
> Actually the "Directly access the it87 registers" option is potentially interesting
> since it will allow using the it87 driver which gives a lot more features.
> 
> I actually wrote a rough outline of how something like this could work here:
> 
> https://bugzilla.kernel.org/show_bug.cgi?id=204807#c47
> 
> Note I'm not sure if that is the right approach, but it definitely is an
> option. It seems that this one might also solve the X470-AORUS-ULTRA-GAMING
> case (https://github.com/t-8ch/linux-gigabyte-wmi-driver/issues/2)
> 
> Hopefully the direct-register ACPI/WMI access methods will also allow
> reading the super-io vendor and product ids so that we can be reasonably
> sure that we are not loading the wrong driver on a board.
> 
> OTOH the WMI-temp method approach may also work on boards where the sensors
> (or some of the sensors) are not supported.
> 
> I'm afraid there is no obviously correct answer here. If you feel like it
> experimenting with the "Directly access the it87 registers" option would certainly
> be interesting IMHO.
> 
> It might be good to get hwmon subsystems maintainer's opinion on this
> before sinking a lot of time into this though (added to the Cc).
> 
> Jean, Guenter,
> 
> Thomas has been working on a WMI driver to expose various motherboard
> temperatures on a gigabyte board where the IO-addresses for the it87 chip
> are reserved by ACPI. We are discussing how best to deal with this, there
> are some ACPI methods to directly access the super-IO registers (with locking
> to protect against other ACPI accesses). This reminded me of an idea I had
> a while ago to solve a similar issue with an other superIO chip, abstract
> the superIO register access-es using some reg_ops struct and allow an ACPI/WMI
> driver to provide alternative reg_ops:
> https://bugzilla.kernel.org/show_bug.cgi?id=204807#c47
> 
> Do you think this is a good idea (or a bad one)? And would something like that
> be acceptable to you ?
> 

The upstream it87 driver is severely out of date. I had an out-of-tree driver
with various improvements which I didn't upstream, first because no one was willing
to review changes and then because it had deviated too much. I pulled it from
public view because I got pounded for not upstreaming it, because people started
demanding support (not asking, demanding) for it, and because Gigabyte stopped
providing datasheets for the more recent ITE chips and it became effectively
unmaintainable.

Some ITE chips have issues which can cause system hangs if accessed directly.
I put some work to remedy that into the non-upstream driver, but that was all
just guesswork. Gigabyte knows about the problem (or so I was told from someone
who has an NDA with them), but I didn't get them or ITE to even acknowledge it
to me. I even had a support case open with Gigabyte for a while, but all I could
get out of them is that they don't support Linux and what I would have to reproduce
the problem with Windows for them to provide assistance (even though, again,
they knew about it).

As for using ACPI locks or WMI to ensure that ACPI leaves the chip alone while
the driver accesses chips directly: That is an option, but it has (at least)
two problems.

First, ACPI access methods are not well documented or standardized. I had tried
to find useful means to do that some time ago, but I gave up because each board
(even from the same vendor) handles locking and accesses differently. We would
end up with lots of board specific code. Coincidentally, that was for ASUS boards
and the nct6775 driver.

Second, access through ACPI is only one of the issues. Turns out there are two
ITE chips on many of the Gigabyte boards nowadays, and the two chips talk to each
other using I2C. My out-of-tree driver tried to remedy that by blocking those
accesses while the driver used the chip, but, again, without Gigabyte / ITE
support this was never a perfect solution, and there was always the risk that
the board ended up hanging because that access was blocked for too long.
Recent ITE chips solve that problem by providing memory mapped access to the
chip registers, but that is only useful if one has a datasheet.

Overall, I don't think it makes much sense trying to make significant changes
to the it87 driver without pulling in all the changes I had made, and without
finding a better fix for the cross-chip access problems. I for sure won't have
time for that (and getting hwmon patches reviewed is still very much an issue).

Having said that, I am of course open to adding WMI/ACPI drivers for the various
boards. Good luck getting support from Gigabyte, though. Or from ASUS, for that
matter.

Guenter