to schedule() or not to schedule() ?

to schedule() or not to schedule() ?

[Kevin Diggs]

Hi,


	I have the following near the top of my cpufreq driver target routine:

while(test_and_set_bit(cf750gxmCfgChangeBit,&cf750gxvStateBits)) {
         /*
          * Someone mucking with our cfg? (I hope it is ok to call
          * schedule() here! - truth is I have no idea what I am doing
          * ... my reasoning is I want to yeild the cpu so whoever is
          * mucking around can finish)
          */
         schedule();
}

This is to prevent bad things from happening if someone is trying to 
change a parameter for the driver via sysfs while the target routine is 
running. Fortunately, because I had a bug where this bit was not getting 
cleared on one of the paths through the target routine ... I now know it 
is not safe to call schedule (it got stuck in there - knocked out my adb 
keyboard! - (I think target is called from a timer that the governor 
sets up ... interrupt context?)).

	How does one very briefly yield the cpu in this context?

kevin

Re: to schedule() or not to schedule() ?

[Chris Friesen]

Kevin Diggs wrote:
> Hi,
> 
> 
>     I have the following near the top of my cpufreq driver target routine:
> 
> while(test_and_set_bit(cf750gxmCfgChangeBit,&cf750gxvStateBits)) {
>         /*
>          * Someone mucking with our cfg? (I hope it is ok to call
>          * schedule() here! - truth is I have no idea what I am doing
>          * ... my reasoning is I want to yeild the cpu so whoever is
>          * mucking around can finish)
>          */
>         schedule();
> }
> 
> This is to prevent bad things from happening if someone is trying to 
> change a parameter for the driver via sysfs while the target routine is 
> running. Fortunately, because I had a bug where this bit was not getting 
> cleared on one of the paths through the target routine ... I now know it 
> is not safe to call schedule (it got stuck in there - knocked out my adb 
> keyboard! - (I think target is called from a timer that the governor 
> sets up ... interrupt context?)).

Is the issue that someone may be in the middle of a multi-stage 
procedure, and you've woken up partway through?

If so, what about simply rescheduling the timer for some short time in 
the future and aborting the current call?

Chris

Re: to schedule() or not to schedule() ?

[Kevin Diggs]

Chris Friesen wrote:
> Kevin Diggs wrote:
> 
>> Hi,
>>
>>
>>     I have the following near the top of my cpufreq driver target 
>> routine:
>>
>> while(test_and_set_bit(cf750gxmCfgChangeBit,&cf750gxvStateBits)) {
>>         /*
>>          * Someone mucking with our cfg? (I hope it is ok to call
>>          * schedule() here! - truth is I have no idea what I am doing
>>          * ... my reasoning is I want to yeild the cpu so whoever is
>>          * mucking around can finish)
>>          */
>>         schedule();
>> }
>>
>> This is to prevent bad things from happening if someone is trying to 
>> change a parameter for the driver via sysfs while the target routine 
>> is running. Fortunately, because I had a bug where this bit was not 
>> getting cleared on one of the paths through the target routine ... I 
>> now know it is not safe to call schedule (it got stuck in there - 
>> knocked out my adb keyboard! - (I think target is called from a timer 
>> that the governor sets up ... interrupt context?)).
> 
> 
> Is the issue that someone may be in the middle of a multi-stage 
> procedure, and you've woken up partway through?
> 
> If so, what about simply rescheduling the timer for some short time in 
> the future and aborting the current call?
> 
> Chris
> 

Chris,

	Thanks for taking the time to reply. The parameter in question modifies 
the frequency table. It is used several times in the target routine. 
I've addressed the issue by making a local copy of the frequency table 
upon entry to the target routine and use that while there. I don't care 
who wins the race.

kevin

Re: to schedule() or not to schedule() ?

[Michael Ellerman]

[-- Attachment #1: Type: text/plain, Size: 2210 bytes --]

On Tue, 2008-08-05 at 12:26 -0700, Kevin Diggs wrote:
> Chris Friesen wrote:
> > Kevin Diggs wrote:

> >>     I have the following near the top of my cpufreq driver target 
> >> routine:
> >>
> >> while(test_and_set_bit(cf750gxmCfgChangeBit,&cf750gxvStateBits)) {
> >>         /*
> >>          * Someone mucking with our cfg? (I hope it is ok to call
> >>          * schedule() here! - truth is I have no idea what I am doing
> >>          * ... my reasoning is I want to yeild the cpu so whoever is
> >>          * mucking around can finish)
> >>          */
> >>         schedule();
> >> }
> >>
> >> This is to prevent bad things from happening if someone is trying to 
> >> change a parameter for the driver via sysfs while the target routine 
> >> is running. Fortunately, because I had a bug where this bit was not 
> >> getting cleared on one of the paths through the target routine ... I 
> >> now know it is not safe to call schedule (it got stuck in there - 
> >> knocked out my adb keyboard! - (I think target is called from a timer 
> >> that the governor sets up ... interrupt context?)).
> > 
> > 
> > Is the issue that someone may be in the middle of a multi-stage 
> > procedure, and you've woken up partway through?
> > 
> > If so, what about simply rescheduling the timer for some short time in 
> > the future and aborting the current call?

> Chris,
> 
> 	Thanks for taking the time to reply. The parameter in question modifies 
> the frequency table. It is used several times in the target routine. 
> I've addressed the issue by making a local copy of the frequency table 
> upon entry to the target routine and use that while there. I don't care 
> who wins the race.

How are you copying the table? Is it an atomic copy? Otherwise you could
just end up copying the table while it's being updated, and you get a
copy of the partially updated table.

Don't you just need a spinlock?

cheers

-- 
Michael Ellerman
OzLabs, IBM Australia Development Lab

wwweb: http://michael.ellerman.id.au
phone: +61 2 6212 1183 (tie line 70 21183)

We do not inherit the earth from our ancestors,
we borrow it from our children. - S.M.A.R.T Person

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: to schedule() or not to schedule() ?

[Kevin Diggs]

Michael Ellerman wrote:
> On Tue, 2008-08-05 at 12:26 -0700, Kevin Diggs wrote:
> 
>>Chris Friesen wrote:
>>
>>>Kevin Diggs wrote:
> 
> 
>>>>    I have the following near the top of my cpufreq driver target 
>>>>routine:
>>>>
>>>>while(test_and_set_bit(cf750gxmCfgChangeBit,&cf750gxvStateBits)) {
>>>>        /*
>>>>         * Someone mucking with our cfg? (I hope it is ok to call
>>>>         * schedule() here! - truth is I have no idea what I am doing
>>>>         * ... my reasoning is I want to yeild the cpu so whoever is
>>>>         * mucking around can finish)
>>>>         */
>>>>        schedule();
>>>>}
>>>>
>>>>This is to prevent bad things from happening if someone is trying to 
>>>>change a parameter for the driver via sysfs while the target routine 
>>>>is running. Fortunately, because I had a bug where this bit was not 
>>>>getting cleared on one of the paths through the target routine ... I 
>>>>now know it is not safe to call schedule (it got stuck in there - 
>>>>knocked out my adb keyboard! - (I think target is called from a timer 
>>>>that the governor sets up ... interrupt context?)).
>>>
>>>
>>>Is the issue that someone may be in the middle of a multi-stage 
>>>procedure, and you've woken up partway through?
>>>
>>>If so, what about simply rescheduling the timer for some short time in 
>>>the future and aborting the current call?
> 
> 
>>Chris,
>>
>>	Thanks for taking the time to reply. The parameter in question modifies 
>>the frequency table. It is used several times in the target routine. 
>>I've addressed the issue by making a local copy of the frequency table 
>>upon entry to the target routine and use that while there. I don't care 
>>who wins the race.
> 
> 
> How are you copying the table? Is it an atomic copy? Otherwise you could
> just end up copying the table while it's being updated, and you get a
> copy of the partially updated table.
> 
> Don't you just need a spinlock?
> 
> cheers
> 
In the initialization routine I create 2 tables. One is a table with all 
the frequencies. The other has just the min and max. The parameter just 
changes a pointer to point to one table or the other. The above 
addressing of the issue should really say "a local copy of the pointer 
to the frequency table".

Thanks for the reply!

For the purpose of learning, there is no direct, correct way to yield 
the cpu when in a timer fired routine, right?

kevin

Re: to schedule() or not to schedule() ?

[Arnd Bergmann]

On Wednesday 06 August 2008, Kevin Diggs wrote:
> For the purpose of learning, there is no direct, correct way to yield 
> the cpu when in a timer fired routine, right?
> 

No, in a timer, you interrupt a totally unrelated thread, so sleeping
would prevent that from running on, as well as preventing other timers
from being run, so it's not an option.

One thing that might work for you would be to re-arm the existing
timer and return from your function, so you get back to it after
a short while.

	Arnd <><