Re: [PATCH 1/2] SOUND: kill gameport bits

[Takashi Iwai <tiwai@suse.de>]

At Sun, 24 Aug 2014 07:07:16 +0200,
Andreas Mohr wrote:
> 
> On Thu, Aug 21, 2014 at 01:29:03PM +0200, Takashi Iwai wrote:
> > I did a quick hack and it seems working on my box.
> > The patch is below.
> 
> Thanks!!
> 
> Further comments below.
> 
> I will be testing this ASAP.
> > +static bool use_ktime = true;
> > +module_param(use_ktime, bool, 0400);
> 
> Towards final commit, should probably add param docs on what may be switched here and why.
> 
> > +
> >  /*
> >   * gameport_mutex protects entire gameport subsystem and is taken
> >   * every time gameport port or driver registrered or unregistered.
> > @@ -76,6 +80,36 @@ static unsigned int get_time_pit(void)
> >  
> >  static int gameport_measure_speed(struct gameport *gameport)
> >  {
> > +	unsigned int i, t, tx;
> > +	u64 t1, t2;
> > +	unsigned long flags;
> > +
> > +	if (gameport_open(gameport, NULL, GAMEPORT_MODE_RAW))
> > +		return 0;
> > +
> > +	tx = ~0;
> > +
> > +	for (i = 0; i < 50; i++) {
> > +		local_irq_save(flags);
> > +		t1 = ktime_get_ns();
> > +		for (t = 0; t < 50; t++)
> > +			gameport_read(gameport);
> > +		t2 = ktime_get_ns();
> > +		local_irq_restore(flags);
> > +		udelay(i * 10);
> > +		if (t2 - t1 < tx)
> > +			tx = t2 - t1;
> 
> This impl is not doing the more complex t3, t2, t1 calculation
> that the PIT impl is doing (likely for the uncommented purpose
> of eliminating timer I/O delay from timing consideration).
> Do/don't ktime/TSC impls better need such an I/O timing correction,
> or are they so fast relative to gameport I/O delays
> that it does not matter? (probably the case for TSC at least).

It's based on x86-64 implementation that doesn't take t3 into
account.  I don't think it doesn't matter so much on the recent
systems, but certainly it can't hurt to measure it, too.

> Oh, and any reason that such a speed calculation remains painfully duplicated
> in both source files? That's possibly done for layering reasons,
> but I'd have to analyze it further.

Yeah, a layer should be one reason.  Another reason is that TSC read
has to be a macro, thus you'd need anyway reimplementation, either
static inline or such.

In my patch, I didn't want to change too much in a shot.  It just adds
the replacement using ktime, that's all.  If you'd like to work on
this further, feel free to do it.

> > +static inline u64 get_time(void)
> > +{
> > +	if (use_ktime) {
> > +		return ktime_get_ns();
> > +	} else {
> > +		unsigned int x;
> > +		GET_TIME(x);
> > +		return x;
> > +	}
> > +}
> 
> It might be useful to have a first commit to introduce these helpers,
> and a second commit to then add ktime support (to keep review code size
> down).

The very purpose of this helper is for ktime.  For TSC, the helper
*is* GET_TIME().  So, splitting commit without introducing ktime
doesn't make much sense.

Nevertheless: did anyone test the patch at all...?


thanks,

Takashi