On Wed, Sep 18, 2013 at 05:23:08PM -0700, Guenter Roeck wrote:
> On Wed, Sep 18, 2013 at 04:32:59PM -0700, Sören Brinkmann wrote:
> > On Wed, Sep 18, 2013 at 04:18:56PM -0700, Joe Perches wrote:
> > > On Wed, 2013-09-18 at 16:09 -0700, Sören Brinkmann wrote:
> > > > On Wed, Sep 18, 2013 at 04:02:41PM -0700, Joe Perches wrote:
> > > > > On Wed, 2013-09-18 at 15:43 -0700, Soren Brinkmann wrote:
> > > > > > Add a driver for SILabs 570, 571, 598, 599 programmable oscillators.
> > > > > > The devices generate low-jitter clock signals and are reprogrammable via
> > > > > > an I2C interface.
> > > > > []
> > > > > > v2:
> > > > > []
> > > > > > - use 10000 as MIN and MAX value in usleep_range
> > > > > []
> > > > > > diff --git a/drivers/clk/clk-si570.c b/drivers/clk/clk-si570.c
> > > > > []
> > > > > > +static int si570_set_frequency(struct clk_si570 *data, unsigned long frequency)
> > > > > > +{
> > > > > []
> > > > > > + /* Applying a new frequency can take up to 10ms */
> > > > > > + usleep_range(10000, 10000);
> > > > >
> > > > > Generally it's nicer to have an actual range for usleep_range.
> > > > Well, as I said in the discussion with Guenther. I'm flexible and nobody
> > > > objected when I said to make both equal. A real range doesn't make sense
> > > > here though, but I don't know what's common practice for cases like
> > > > this.
> > >
> > > udelay is normal, but I guess you don't need atomic context.
> > After checkpatch correcting me a few times I went with what
> > Documentation/timers/timers-howto.txt suggests. But yes, then we have
> > this situation, that I want to sleep 10ms, but not longer using a
> > *_range function. I guess it is very application specific whether a
> > longer delay here is acceptable or not.
> >
> You really want to sleep and not call udelay for 10ms. The idea behind usleep_range
> is that you give the kernel some slack. In this case, you could for example make it
> 10-12 ms. That doesn't make much difference for the driver, but it might save a
> timer interrupt in the kernel because it might be able to coalesce more than one
> event. After all, it doesn't have to be _exactly_ 10 ms, which is what you are
> claiming with the fixed number. Prior to usleep_range, you would have happily
> called msleep(10) without realizing that it might sleep up to 20 ms on you.
> Keep that in mind ...
>
> > You're right. I'll add a delay there as well. The 'rang' question
> > applies here as well.
> >
> Same thing, really. You could make it 100-200uS. That doesn't make much
> difference for this driver, but it might make a difference for overall
> performance, especially if everyone is playing nicely.
>
Okay, so I'll use a real range. 10 - 12 ms for big frequency changes and
100 - 200 us for small ones, as Guenther suggests. Does that sound okay?
Thanks,
Sören
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
