Re: [PATCH v3] PM / QoS: add pm_qos_update_request_timeout API

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Monday, March 26, 2012, MyungJoo Ham wrote:
> On Mon, Mar 26, 2012 at 12:02 PM, mark gross <markgross@thegnar.org> wrote:
> > On Mon, Mar 26, 2012 at 10:41:15AM +0900, MyungJoo Ham wrote:
> >> On Sun, Mar 25, 2012 at 1:35 AM, mark gross <markgross@thegnar.org> wrote:
> >> > I apologize for the lat replay and admit that I was probably wrong to
> >> > oppose the idea of time out pm_qos requests.  (last week we bumped into
> >> > a need for them and now I get it.)
> >> >
> >> >
> >> > On Wed, Mar 07, 2012 at 02:06:18PM +0900, MyungJoo Ham wrote:
> >> >> The new API, pm_qos_update_request_timeout() is to provide a timeout
> >> >> with pm_qos_update_request.
> >> >>
> >> >> For example, pm_qos_update_request_timeout(req, 100, 1000), means that
> >> >> QoS request on req with value 100 will be active for 1000 jiffies.
> >> >> After 1000 jiffies, the QoS request thru req is rolled back to the
> >> >> request status when pm_qos_update_request_timeout() was called. If there
> >> >> were another pm_qos_update_request(req, x) during the 1000 jiffies, this
> >> >> new request with value x will override as this is another request on the
> >> >> same req handle. A new request on the same req handle will always
> >> >> override the previous request whether it is the conventional request or
> >> >> it is the new timeout request.
> >> >>
> >> >> Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>
> >> >> Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
> >> []
> >> >> @@ -77,6 +79,8 @@ void pm_qos_add_request(struct pm_qos_request *req, int pm_qos_class,
> >> >>                       s32 value);
> >> >>  void pm_qos_update_request(struct pm_qos_request *req,
> >> >>                          s32 new_value);
> >> >> +void pm_qos_update_request_timeout(struct pm_qos_request *req,
> >> >> +                                s32 new_value, unsigned long timeout_ms);
> >> > is ms the right units?  could we ever need us?
> >> >
> >>
> >> Because jiffies are used for scheduling tasks, I thought ms should be
> >> fine and having some devices running fast for some msecs longer won't
> >> hurt. However, do you expect scheduling tasks or jiffies may use usecs
> >> later? I don't mind using usecs instead of msecs here; thus, I'll
> >> update this to use usecs. I'll resend patchset soon.
> >>
> >
> > I am just asking a question.  I'm not sure if us or ms are the better
> > units off the top of my head.
> >
> > --mark
> >
> 
> 
> For the current structure of Linux (scheduling task, jiffies), I
> thought that none of the two is better than the another because the
> duration should not be so long and the jiffies are larger than 1 ms.
> 
> However, it turns out to be not true: some uses jiffies < 1ms (alpha),
> some uses jiffies not cleanly dividable with msecs (omap). Thus, it
> appears that usecs is better. Thus, the parameter should be usecs
> rather than msecs.

Yes, usecs are generally better, although you may consider using nsecs too.

Thanks,
Rafael