on 11/07/2011 19:07 Vitaly Magerya said the following:
> Andriy Gapon <a...@freebsd.org> wrote:
>> on 06/07/2011 22:20 Vitaly Magerya said the following:
>>> --- acpi_cpu.c.orig     2011-07-05 19:50:31.000000000 +0000
>>> +++ acpi_cpu.c  2011-07-06 17:23:16.000000000 +0000
>>> @@ -1194,7 +1194,7 @@
>>>      if (strlen(state) < 2 || toupper(state[0]) != 'C')
>>>         return (EINVAL);
>>>      val = (int) strtol(state + 1, NULL, 10) - 1;
>>> -    if (val < 0 || val > cpu_cx_count - 1)
>>> +    if (val < 0)
>>>         return (EINVAL);
>>>      cpu_cx_lowest = val;
>>
>> This change is a little bit more intrusive than I would like.
>> There are some things about cpu_cx_lowest handling in the code that make me 
>> a bit
>> unsure if this change is completely safe.
> 
> Can you elaborate? From my reading, the only place cpu_cx_lowest
> is used is in acpi_cpu_notify, where sc->cpu_cx_lowest is optionally
> increased to min(cpu_cx_lowest, sc->cpu_cx_count - 1), which should
> be safe in any situation.

This is exactly the place that I am concerned about.
Probably my mind is clouded but I can not understand why 
acpi_cpu_set_cx_lowest()
call is under the condition:
    if (sc->cpu_cx_lowest < cpu_cx_lowest)
        acpi_cpu_set_cx_lowest(sc, min(cpu_cx_lowest, sc->cpu_cx_count - 1));

If you or someone else can explain to me why that condition is there...

> Also note that we currently do not update cpu_cx_lowest immediately
> when the number of available states changes (only devd+power_profile
> does that). For example, if I kill devd and plug the power cord
> off, cpu_cx_lowest remains at C3, even though only C2 is reported.
> This is why the above patch shouldn't introduce situations we don't
> already have.

Yes, quite a good point.
Although I am not sure yet if what you describe is not a bug that should be 
fixed.

>> I suspect that there could be problems
>> on systems where number Cx states becomes smaller after some events (e.g. AC 
>> connection).
> 
> I have such a system; if there are situations you'd like me to test,
> I can do so (so far it looks good).

I am not exactly sure what to look for...
Perhaps something like this (if your system would allow it):
- place the system in a state where at least C3 is supported
- set global cx_lowest to C3
- set per-CPU cx_lowest for one CPU to C2
- place a system in a state where only C1 is supported

This testcase is only tangentially related to your proposed change.  It's more
about that code that I don't understand.

-- 
Andriy Gapon
_______________________________________________
freebsd-acpi@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-acpi
To unsubscribe, send any mail to "freebsd-acpi-unsubscr...@freebsd.org"

Reply via email to