Re: unparseable, undocumented /sys/class/drm/.../pstate

From: Ilia Mirkin
Date: Mon Jun 23 2014 - 14:05:21 EST


On Mon, Jun 23, 2014 at 2:00 PM, Martin Peres <martin.peres@xxxxxxx> wrote:
> Le 23/06/2014 19:56, Ilia Mirkin a Ãcrit :
>
>> On Mon, Jun 23, 2014 at 1:46 PM, Martin Peres <martin.peres@xxxxxxx>
>> wrote:
>>>
>>> Le 23/06/2014 18:40, Ilia Mirkin a Ãcrit :
>>>>
>>>>
>>>> On Mon, Jun 23, 2014 at 12:36 PM, Greg KH <greg@xxxxxxxxx> wrote:
>>>>>
>>>>>
>>>>> On Mon, Jun 23, 2014 at 12:18:51PM -0400, Ilia Mirkin wrote:
>>>>> A list of valid "values" that a file can be in is fine if you just then
>>>>> write one value back to that file. That's the one exception, but a
>>>>> minor one given the huge number of sysfs files. Other than that, if
>>>>> you
>>>>
>>>>
>>>>
>>>> Which is pretty much what the pstate file is. Would it make things
>>>> better if we removed the descriptive info while leaving the pstate
>>>> file in place?
>>>
>>>
>>>
>>> This means we should also create a new sysfs file per performance level
>>> too,
>>> right? Is there another way for a driver to expose a list in sysfs?
>>>
>>> Since NVIDIA gives different names to performance levels depending on the
>>> card family, we may need to abstract the name away in order to provide
>>> some
>>> consistency and make listing performance levels easier from a program
>>> (may
>>> it use readdir() or stat()).
>>>
>>> Moving the file to debugfs would "fix" the one-value-per-file rule but it
>>> would also require users to mount debugfs at boot time in order to write
>>> the
>>> default configuration they want for PM instead of just changing
>>> /etc/sysctl.d/nouveau.conf... On the other hand, I'm not sure we can
>>> commit
>>> on having a stable ABI on the way we display clocks (unless people take
>>> them
>>> as a single value and do not try to parse them) as new hardware will
>>> alter
>>> the semantics of each clock domain, if not drop/split some of them!
>>>
>>> Whatever we do, it doesn't look like we can find a nice solution that
>>> fits
>>> every use cases unless we write a userspace program to access this data,
>>> but
>>> this seems highly overkill...
>>
>>
>> I was thinking just having the list of level ids in the pstate file,
>> and then stick the current file into debugfs. That way people retain
>> the ability to see things, as well as use pstate directly for a
>> configured system.
>
>
> In this case, would we still use the * to indicate the current perflvl?

That would only be in debugfs. pstate would just list the available
levels and let you set one. (or 'auto', as you point out below)

>
> Also, are we supposed to output the current perflvl or the current
> configuration in use? Right now, we configure it to either auto (WIP),
> perflvl X at all time or perflvl X when on battery and Y when on sector.
> However, when we read pstate, we only get the current perflvl if my memory
> serves me right. Maybe we should create a r-o file that outputs the current
> perflvl and keep pstate for storing the configuration.

Yeah, we could do something like that... ugh, the battery/ac stuff is
a complication. Unclear whether that belongs in the kernel in the
first place... but I guess other drivers do it?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/