[PATCH] Real-Time Preemption, using msecs_to_jiffies() instead of HZ

[PATCH] Real-Time Preemption, using msecs_to_jiffies() instead of HZ

[Luca Falavigna]

Hi,

I was looking at kernel/softlookup.c when I noticed you used HZ in order to get
a 10-second delay:

void softlockup_tick(struct pt_regs *regs)
{
	...
	if (time_after(jiffies, timestamp + 10*HZ)) {
	...
}

I created this small patch (built against version 2.6.12-rc6-V0.7.48-05) which
does use of msecs_to_jiffies() to get a correct behaviour with every platform.
Similarly I modified function watchdog and kernel/irq/autoprobe.c file
(probe_irq_on function).

Here is the patch:

--- ./rtp-2.6.12-rc6-V0.7.48-05.orig	2005-06-10 16:27:48.000000000 +0000
+++ ./rtp-2.6.12-rc6-V0.7.48-05		2005-06-10 16:30:18.000000000 +0000
@@ -7425,7 +7425,7 @@
 +	/*
 +	 * Wait for longstanding interrupts to trigger, 20 msec delay:
 +	 */
-+	msleep(HZ/50);
++	msleep(msecs_to_jiffies(20));

  	/*
  	 * enable any unassigned irqs
@@ -7452,7 +7452,7 @@
  	 */
 -	for (delay = jiffies + HZ/10; time_after(delay, jiffies); )
 -		/* about 100ms delay */ barrier();
-+	msleep(HZ/10);
++	msleep(msecs_to_jiffies(100));

  	/*
  	 * Now filter out any obviously spurious interrupts
@@ -10712,7 +10712,7 @@
 +	if (did_panic)
 +		return;
 +
-+	if (time_after(jiffies, timestamp + 10*HZ)) {
++	if (time_after(jiffies, timestamp + msecs_to_jiffies(10000))) {
 +		per_cpu(print_timestamp, this_cpu) = timestamp;
 +
 +		spin_lock(&print_lock);
@@ -10748,7 +10748,7 @@
 +	 */
 +	while (!kthread_should_stop()) {
 +		set_current_state(TASK_INTERRUPTIBLE);
-+		msleep_interruptible(HZ);
++		msleep_interruptible(msecs_to_jiffies(1000));
 +		touch_softlockup_watchdog();
 +	}
 +	__set_current_state(TASK_RUNNING);



Regards,

-- 
					Luca

Re: [PATCH] Real-Time Preemption, using msecs_to_jiffies() instead of HZ

[Ingo Molnar]

* Luca Falavigna <dktrkranz@gmail.com> wrote:

> I was looking at kernel/softlookup.c when I noticed you used HZ in order to get
> a 10-second delay:
> 
> void softlockup_tick(struct pt_regs *regs)
> {
> 	...
> 	if (time_after(jiffies, timestamp + 10*HZ)) {
> 	...
> }

oops, indeed. (i've also forwarded your patch to akpm, as the softlockup 
patch is included in -mm too)

> I created this small patch (built against version 
> 2.6.12-rc6-V0.7.48-05) which does use of msecs_to_jiffies() to get a 
> correct behaviour with every platform. Similarly I modified function 
> watchdog and kernel/irq/autoprobe.c file (probe_irq_on function).
> 
> Here is the patch:

> ++	msleep(msecs_to_jiffies(20));
> ++	msleep(msecs_to_jiffies(100));
> ++		msleep_interruptible(msecs_to_jiffies(1000));

actually, this should be msleep(20/100/1000). I've fixed these in my 
tree.

	Ingo

Re: [PATCH] Real-Time Preemption, using msecs_to_jiffies() instead of HZ

[Luca Falavigna]

Ingo Molnar ha scritto:
> oops, indeed. (i've also forwarded your patch to akpm, as the softlockup 
> patch is included in -mm too)

OK, thank you!


I also found another couple of those in drivers/char/rtc.c file.
Here is a quick fix:

--- ./rtp-2.6.12-rc6-V0.7.48-17.orig	2005-06-13 01:26:53.000000000 +0000
+++ ./rtp-2.6.12-rc6-V0.7.48-17		2005-06-13 01:58:13.000000000 +0000
@@ -17044,7 +17044,7 @@
 -			barrier();
 -			cpu_relax();
 -		}
-+		msleep(2*HZ/100);
++		msleep(20);
  	
  	spin_lock_irq(&rtc_lock);
  	year = CMOS_READ(RTC_YEAR);
@@ -17102,7 +17102,7 @@
 -			barrier();
 -			cpu_relax();
 -		}
-+		msleep(2*HZ/100);
++		msleep(20);

  	/*
  	 * Only the values that we read from the RTC are set. We leave

Regards,
-- 
					Luca