Re: [ath9k-devel] [PATCH 1/3] ath9k: Fix build error on ARM

[Joe Perches <joe@perches.com>]

On Wed, 2014-02-05 at 14:21 +0000, Russell King - ARM Linux wrote:
> On Wed, Feb 05, 2014 at 06:03:13AM -0800, Joe Perches wrote:
> > On Wed, 2014-02-05 at 13:39 +0000, Russell King - ARM Linux wrote:
> > > On Wed, Feb 05, 2014 at 05:04:54AM -0800, Joe Perches wrote:
> > > > On Wed, 2014-02-05 at 12:41 +0000, Russell King - ARM Linux wrote:
> > > > > On Wed, Feb 05, 2014 at 04:32:46AM -0800, Joe Perches wrote:
> > > > > > Apparently, people just convert stupidly large udelay()s
> > > > > > to mdelay and not be bothered.
> > > > > 
> > > > > And that's the correct answer.  Having udelay(10000) rather than mdelay(10)
> > > > > is a sign that they weren't paying that much attention when writing the
> > > > > code.
> > > > 
> > > > Not really.
> > []
> > > > It's not so much not paying attention as not
> > > > knowing ARM is broken for large udelay().
> > > 
> > > And now read my suggestion about how to avoid the "not knowing" problem. :)
> > 
> > I'd read it already.  I didn't and don't disagree.
> 
> Please explain /why/ you don't agree.

Please reread what I wrote.

We agree a lot more than you seem to think.

> > I still think adding a #warning on large static udelay()s
> > would be sensible.  Maybe adding another option like
> > #define UDELAY_TOO_BIG_I_KNOW_ALREADY_DONT_BOTHER_ME
> > guard to avoid seeing the #warning when there's no other
> > option.
> 
> How does that help?  It's /not/ a warning situation - if you ask for
> udelay(10000) on ARM, you will /not/ get a 10ms delay.  You'll instead
> get something much shorter because the arithmetic will overflow.

> Now, you could argue that maybe ARM's udelay() implementation should
> know about this and implement a loop around the underlying udelay()
> implementation to achieve it,

I suggested something like that was possible.

> and I might agree with you - but I
> don't for one very simple reason: we /already/ have such an
> implementation.  It's called mdelay():

I think mdelay should be for the case where the range
is too big for a udelay.  I think arm's 11 bit range
is unfortunately small.

I think we agree be nice to get a generic compiler
#warning when a __builtin_constant_p value is too large
and a ratelimited dmesg/warning for the runtime case.