Skip to contentSkip to conversation

Is 64bit LVTT screwed

Q Cyrill Gorcunov 2 participants 4 messages Jul 2, 2008 – Jul 8, 2008 x86

lore
Share & tools
lore
#0 lore raw

Hi Maciej,

while I'm in path of unification apic code I found a bit strange code snippet

apic_32.c
---------
#define APIC_DIVISOR 16
static void __setup_APIC_LVTT(unsigned int clocks, int oneshot, int irqen)

{
	...
	if (!oneshot)
		apic_write_around(APIC_TMICT, clocks / APIC_DIVISOR);

}

apic_64.c
---------
static void __setup_APIC_LVTT(unsigned int clocks, int oneshot, int irqen)

{
        ...
	if (!oneshot)
		apic_write_around(APIC_TMICT, clocks);
}

but in both cases we use "divide by 16" in divide register. The only explanation I imagine - for 64bit mode we are required to 'stuck' for a bit longer (by 16 times longer to be precise). Am I right? Or there is another reason why we dont use APIC_DIVISOR here. Actually, as I see it not fair to a caller. For 64bit mode APIC timer is requested to count 250000000 ticks but in real it will count 250000000 * 16. Not sure who is right there. I think the better would be to use 4000000000 and APIC_DIVISOR in 64bit mode. How do you think?

- Cyrill -

#1 lore raw
Quoted 23 lines · Cyrill Gorcunov

while I'm in path of unification apic code I found a bit strange code snippet

apic_32.c
---------
#define APIC_DIVISOR 16
static void __setup_APIC_LVTT(unsigned int clocks, int oneshot, int irqen)

{
	...
	if (!oneshot)
		apic_write_around(APIC_TMICT, clocks / APIC_DIVISOR);

}

apic_64.c
---------
static void __setup_APIC_LVTT(unsigned int clocks, int oneshot, int irqen)

{
        ...
	if (!oneshot)
		apic_write_around(APIC_TMICT, clocks);
}
Quoted 8 lines

but in both cases we use "divide by 16" in divide register. The only explanation I imagine - for 64bit mode we are required to 'stuck' for a bit longer (by 16 times longer to be precise). Am I right? Or there is another reason why we dont use APIC_DIVISOR here. Actually, as I see it not fair to a caller. For 64bit mode APIC timer is requested to count 250000000 ticks but in real it will count 250000000 * 16. Not sure who is right there. I think the better would be to use 4000000000 and APIC_DIVISOR in 64bit mode. How do you think?

The APIC clock varies across systems so the timer setup includes
calibration. Which means both variations end up with the correct interrupt rate probably, except using different divisors. We used to print the APIC clock rate and that would be off with one of the above, but I don't we do this anymore. We had a similar problem with the 82489DX where the prescaler was bypassed altogether resulting in an incorrect clock rate printed, but the rate of timer interrupt was correctly calibrated regardless.

This is of course merely an explanation why the code you quoted does not
explode spectacularly. It does not make it correct. If the prescaler is indeed set to 16 for 64-bit systems, then APIC_DIVISOR should obviously be used in the calculation above too. And if not, then I think the prescaler should be set consistently across systems and then the setting reflected in the calculations accordingly. Please note that only values between 2 and 16 are supported in a uniform way across all the APIC models and using low values risks an overflow as clock rates are in the GHz range these days already. The value of 16 seems a reasonable one.

Maciej

#2 lore raw

On Tue, Jul 8, 2008 at 2:08 AM, Maciej W. Rozycki <[email protected]> wrote:
[...]

Quoted 29 lines
Quoted 9 lines

but in both cases we use "divide by 16" in divide register. The only explanation I imagine - for 64bit mode we are required to 'stuck' for a bit longer (by 16 times longer to be precise). Am I right? Or there is another reason why we dont use APIC_DIVISOR here. Actually, as I see it not fair to a caller. For 64bit mode APIC timer is requested to count 250000000 ticks but in real it will count 250000000 * 16. Not sure who is right there. I think the better would be to use 4000000000 and APIC_DIVISOR in 64bit mode. How do you think?

The APIC clock varies across systems so the timer setup includes
calibration. Which means both variations end up with the correct interrupt rate probably, except using different divisors. We used to print the APIC clock rate and that would be off with one of the above, but I don't we do this anymore. We had a similar problem with the 82489DX where the prescaler was bypassed altogether resulting in an incorrect clock rate printed, but the rate of timer interrupt was correctly calibrated regardless.

This is of course merely an explanation why the code you quoted does not
explode spectacularly. It does not make it correct. If the prescaler is indeed set to 16 for 64-bit systems, then APIC_DIVISOR should obviously be used in the calculation above too. And if not, then I think the prescaler should be set consistently across systems and then the setting reflected in the calculations accordingly. Please note that only values between 2 and 16 are supported in a uniform way across all the APIC models and using low values risks an overflow as clock rates are in the GHz range these days already. The value of 16 seems a reasonable one.

Quoted 2 lines

Maciej

Thanks, Maciej, for explanation. Unfortunelly I'm quite busy now so I stopped merging/research on apic code, as only get spare time - will let you know my results.

- Cyrill -

#3 lore raw
In reply to Maciej W. Rozycki

[Maciej W. Rozycki - Mon, Jul 07, 2008 at 11:08:57PM +0100] [...]

| > 
| > but in both cases we use "divide by 16" in divide register. The only
| > explanation I imagine - for 64bit mode we are required to 'stuck'
| > for a bit longer (by 16 times longer to be precise). Am I right?
| > Or there is another reason why we dont use APIC_DIVISOR here. Actually,
| > as I see it not fair to a caller. For 64bit mode APIC timer is requested
| > to count 250000000 ticks but in real it will count 250000000 * 16.
| > Not sure who is right there. I think the better would be to
| > use 4000000000 and APIC_DIVISOR in 64bit mode. How do you think?
| 
|  The APIC clock varies across systems so the timer setup includes
| calibration.  Which means both variations end up with the correct
| interrupt rate probably, except using different divisors.  We used to
| print the APIC clock rate and that would be off with one of the above, but
| I don't we do this anymore.  We had a similar problem with the 82489DX
| where the prescaler was bypassed altogether resulting in an incorrect 
| clock rate printed, but the rate of timer interrupt was correctly 
| calibrated regardless.
| 
|  This is of course merely an explanation why the code you quoted does not
| explode spectacularly.  It does not make it correct.  If the prescaler is
| indeed set to 16 for 64-bit systems, then APIC_DIVISOR should obviously be
| used in the calculation above too.  And if not, then I think the prescaler 
| should be set consistently across systems and then the setting reflected 
| in the calculations accordingly.  Please note that only values between 2 
| and 16 are supported in a uniform way across all the APIC models and using 
| low values risks an overflow as clock rates are in the GHz range these 
| days already.  The value of 16 seems a reasonable one.
| 
|   Maciej
|

Maciej, I think the right soulution is the next:

arch/x86/kernel/apic_64.c +5 -2
Index: linux-2.6.git/arch/x86/kernel/apic_64.c
===================================================================
@@ -165,6 +165,9 @@ int lapic_get_maxlvt(void)
 	return maxlvt;
 }
 
+/* Clock divisor is set to 16 */
+#define APIC_DIVISOR 16
+
 /*
  * This function sets up the local APIC timer, with a timeout of
  * 'clocks' APIC bus clock. During calibration we actually call
@@ -197,7 +200,7 @@ static void __setup_APIC_LVTT(unsigned i
 				| APIC_TDR_DIV_16);
 
 	if (!oneshot)
-		apic_write(APIC_TMICT, clocks);
+		apic_write(APIC_TMICT, clocks/APIC_DIVISOR);
 }
 
 /*
@@ -329,7 +332,7 @@ static void __init calibrate_APIC_clock(
 	 *
 	 * No interrupt enable !
 	 */
-	__setup_APIC_LVTT(250000000, 0, 0);
+	__setup_APIC_LVTT(4000000000, 0, 0);
 
 	apic_start = apic_read(APIC_TMCCT);
 #ifdef CONFIG_X86_PM_TIMER

So we remain behaviour the same (ie calibrating duration remains the same) but it has touched lapic_timer_setup. Will take a look into this more closer soon.

- Cyrill -