Re: [RFC PATCH v2 12/14] x86/watchdog/hardlockup/hpet: Determine if HPET timer caused NMI

[Ricardo Neri]

On Tue, Apr 09, 2019 at 01:28:17PM +0200, Peter Zijlstra wrote:
> On Wed, Feb 27, 2019 at 08:05:16AM -0800, Ricardo Neri wrote:
> > @@ -62,7 +67,18 @@ static inline void set_comparator(struct hpet_hld_data *hdata,
> >  static void kick_timer(struct hpet_hld_data *hdata, bool force)
> >  {
> >  	bool kick_needed = force || !(hdata->flags & HPET_DEV_PERI_CAP);
> > -	unsigned long new_compare, count;
> > +	unsigned long tsc_curr, tsc_delta, new_compare, count;
> > +
> > +	/* Start obtaining the current TSC and HPET counts. */
> > +	tsc_curr = rdtsc();
> > +
> > +	if (kick_needed)
> > +		count = get_count();
> > +
> > +	tsc_delta = (unsigned long)watchdog_thresh * (unsigned long)tsc_khz
> > +		    * 1000L;
> > +	hdata->tsc_next = tsc_curr + tsc_delta;
> > +	hdata->tsc_next_error = tsc_delta >> 6;
> 
> What do we need a per hld_data tsc_next_error for? It is basically a
> global 'constant'.
> 

This is true. I thought I'd keep all the needed variables in a single
struct to make the code more readable. I guess, I did not achieve that
goal. I'll put it as a static global variable.

> >  	/*
> >  	 * Update the comparator in increments of watch_thresh seconds relative
> > @@ -74,8 +90,6 @@ static void kick_timer(struct hpet_hld_data *hdata, bool force)
> >  	 */
> >  
> >  	if (kick_needed) {
> > -		count = get_count();
> > -
> >  		new_compare = count + watchdog_thresh * hdata->ticks_per_second;
> >  
> >  		set_comparator(hdata, new_compare);
> > @@ -147,6 +161,14 @@ static void set_periodic(struct hpet_hld_data *hdata)
> >   */
> >  static bool is_hpet_wdt_interrupt(struct hpet_hld_data *hdata)
> >  {
> > +	if (smp_processor_id() == hdata->handling_cpu) {
> > +		unsigned long tsc_curr;
> 
> TSC is u64

In x86_64, isn't u64 an unsigned long? Do you mean to consider the
32-bit case?

> 
> > +
> > +		tsc_curr = rdtsc();
> > +		if (abs(tsc_curr - hdata->tsc_next) < hdata->tsc_next_error)
> > +			return true;
> 
> You can write that as:
> 
> 		(tsc_curr - hdata->tsc_next) + tsc_error < 2*tsc_error
> 
> which doesn't contain any branches what so ever.
> 

Sure, I'll add this change.

Thanks and BR,
Ricardo