Skip to contentSkip to conversation

fix #endif misplacement

PATCH In review Ricardo Nabinger Sanchez 6 participants 11 messages Nov 28, 2003

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fix #endif misplacement

This patch fixes an #endif misplacement, which leads to dead code in sched_clock() in arch/i386/kernel/timers/timer_tsc.c, due to a return outside the ifdef/endif.

Please consider applying, as sched_clock() apparently does not behave as expected. Patched against 2.6.0-test11.

Regards.

Signature · 24 lines
Ricardo Nabinger Sanchez
GNU/Linux #140696 [http://counter.li.org]
Slackware Linux

  Warning: 
    Trespassers will be shot.
    Survivors will be shot again.



diff -urN linux-2.6.0-test11/arch/i386/kernel/timers/timer_tsc.c
linux-2.6.0-test11-sched_clock/arch/i386/kernel/timers/timer_tsc.c
--- linux-2.6.0-test11/arch/i386/kernel/timers/timer_tsc.c	2003-11-26 18:44:45.000000000 -0200
+++ linux-2.6.0-test11-sched_clock/arch/i386/kernel/timers/timer_tsc.c	2003-11-28 12:58:59.000000000 -0200
@@ -140,8 +140,8 @@
 	 */
 #ifndef CONFIG_NUMA
 	if (!use_tsc)
-#endif
 		return (unsigned long long)jiffies * (1000000000 / HZ);
+#endif
 
 	/* Read the Time Stamp Counter */
 	rdtscll(this_offset);
#1 lore raw
Quoted 3 lines · Ricardo Nabinger Sanchez

This patch fixes an #endif misplacement, which leads to dead code in sched_clock() in arch/i386/kernel/timers/timer_tsc.c, due to a return outside the ifdef/endif.

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.
Look at the comment just above those lines.

Tim

Quoted 12 lines
arch/i386/kernel/timers/timer_tsc.c +1 -1
@@ -140,8 +140,8 @@
 	 */
 #ifndef CONFIG_NUMA
 	if (!use_tsc)
-#endif
 		return (unsigned long long)jiffies * (1000000000 / HZ);
+#endif
 
 	/* Read the Time Stamp Counter */
 	rdtscll(this_offset);
#2 lore raw

Hi Tim,

Quoted 1 line · Tim Schmielau

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.

The patch didn't hurt this.

Quoted 1 line

Look at the comment just above those lines.

The patch doesn't uses jiffies indiscriminately: Only if we're on a
NUMA system with !use_tsc.

Otherwise (on x86 SMP, for example) we use rdtsc...which seems The
Right Thing(tm). Hece move the #endif a bit down.

Cheers

Felipe

#3 lore raw

Hi,

Actually, the original code seems to be ok:

#ifndef CONFIG_NUMA
        if (!use_tsc)
#endif
        return (unsigned long long)jiffies * (1000000000 / HZ);

That is: on x86 we'll get into "#ifndef CONFIG_NUMA", and "if (!use_tsc)" will be called, which should be the expected behaviour.

Maybe in your case the use_tsc flag was being set to 0 (bug in detection code / unsupported feature by the processor?), leading your code to ignore the TSC and return the current time based on the current jiffies instead.

Lucas

Quoted 25 lines · Felipe W Damasio

Hi Tim,

Quoted 2 lines · Tim Schmielau

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.

The patch didn't hurt this.

Quoted 1 line

Look at the comment just above those lines.

The patch doesn't uses jiffies indiscriminately: Only if we're on a
NUMA system with !use_tsc.

Otherwise (on x86 SMP, for example) we use rdtsc...which seems The
Right Thing(tm). Hece move the #endif a bit down.

Cheers

Felipe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#4 lore raw
In reply to Felipe W Damasio
Quoted 9 lines · Felipe W Damasio

Hi Tim,

Quoted 1 line · Tim Schmielau

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.

The patch didn't hurt this.

Quoted 8 lines
Quoted 1 line

Look at the comment just above those lines.

The patch doesn't uses jiffies indiscriminately: Only if we're on
a NUMA system with !use_tsc.

Otherwise (on x86 SMP, for example) we use rdtsc...which seems The
Right Thing(tm). Hece move the #endif a bit down.

The ifdef isn't pretty, but its performance critical code, its easy to understand, and there is a big comment above it. I think its OK the way it is. Not that you would ever notice any difference probably.

#5 lore raw

Hi Nick,

Quoted 3 lines · Nick Piggin

The ifdef isn't pretty, but its performance critical code, its easy to understand, and there is a big comment above it. I think its OK the way it is. Not that you would ever notice any difference probably.

You're right. As Lucas already pointed out, the ifdef CONFIG_NUMA is
actually an ifndef...

Like myself, I think Ricardo overlooked this :)

Cheers,

Felipe

#6 lore raw

Quoting Felipe W Damasio <[email protected]>
Sent on Fri, 28 Nov 2003 15:12:20 -0200

Quoted 10 lines

Hi Nick,

Quoted 3 lines · Nick Piggin

The ifdef isn't pretty, but its performance critical code, its easy to understand, and there is a big comment above it. I think its OK the way it is. Not that you would ever notice any difference probably.

You're right. As Lucas already pointed out, the ifdef CONFIG_NUMA is
actually an ifndef...
Like myself, I think Ricardo overlooked this :)

Oops! Completely overlooked this. Thanks for correcting me :)

Signature · 7 lines
Ricardo Nabinger Sanchez
GNU/Linux #140696 [http://counter.li.org]
Slackware Linux

  Warning: 
    Trespassers will be shot.
    Survivors will be shot again.
#7 lore raw
In reply to Tim Schmielau
Quoted 9 lines · Ricardo Nabinger Sanchez
Quoted 4 lines

This patch fixes an #endif misplacement, which leads to dead code in sched_clock() in arch/i386/kernel/timers/timer_tsc.c, due to a return outside the ifdef/endif.

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.
Look at the comment just above those lines.

HOWEVER, we seem to have some jiffies wrap bugs in this file.

Tim

arch/i386/kernel/timers/timer_tsc.c renamed from arch/i386/kernel/timers/timer_tsc.c.orig +6 -5
@@ -30,7 +30,6 @@
 int tsc_disable __initdata = 0;
 
 extern spinlock_t i8253_lock;
-extern volatile unsigned long jiffies;
 
 static int use_tsc;
 /* Number of usecs that the last interrupt was delayed */
@@ -141,7 +140,7 @@
 #ifndef CONFIG_NUMA
 	if (!use_tsc)
 #endif
-		return (unsigned long long)jiffies * (1000000000 / HZ);
+		return (unsigned long long)get_jiffies_64() * (1000000000 / HZ);
 
 	/* Read the Time Stamp Counter */
 	rdtscll(this_offset);
@@ -215,7 +214,8 @@
 	lost = delta/(1000000/HZ);
 	delay = delta%(1000000/HZ);
 	if (lost >= 2) {
-		jiffies += lost-1;
+		/* only called under xtime_lock */
+		jiffies_64 += lost-1;
 
 		/* sanity check to ensure we're not always losing ticks */
 		if (lost_count++ > 100) {
@@ -241,7 +241,7 @@
 	 * usec delta is > 90% # of usecs/tick)
 	 */
 	if (lost && abs(delay - delay_at_last_interrupt) > (900000/HZ))
-		jiffies++;
+		jiffies_64++;
 }
 
 static void delay_tsc(unsigned long loops)
@@ -283,7 +283,8 @@
 	offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
 	if (unlikely(((offset - hpet_last) > hpet_tick) && (hpet_last != 0))) {
 		int lost_ticks = (offset - hpet_last) / hpet_tick;
-		jiffies += lost_ticks;
+		/* only called under xtime_lock */
+		jiffies_64 += lost_ticks;
 	}
 	hpet_last = hpet_current;
 
#8 lore raw

PATCH 1/1 another jiffies wrap bug

Quoted 1 line

HOWEVER, we seem to have some jiffies wrap bugs in this file.

Should have checked the whole directory first, the chance of cut'n paste bugs was just too high...

Tim

arch/i386/kernel/timers/timer_hpet.c renamed from arch/i386/kernel/timers/timer_hpet.c.orig +2 -1
@@ -108,7 +108,8 @@
 	offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
 	if (unlikely(((offset - hpet_last) > hpet_tick) && (hpet_last != 0))) {
 		int lost_ticks = (offset - hpet_last) / hpet_tick;
-		jiffies += lost_ticks;
+		/* only called under xtime_lock */
+		jiffies_64 += lost_ticks;
 	}
 	hpet_last = offset;
 
#9 lore raw
In reply to Tim Schmielau

Quoting Tim Schmielau <[email protected]> Sent on Fri, 28 Nov 2003 17:34:49 +0100 (CET)

Quoted 7 lines
Quoted 3 lines

This patch fixes an #endif misplacement, which leads to dead code in sched_clock() in arch/i386/kernel/timers/timer_tsc.c, due to a return outside the ifdef/endif.

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.
Look at the comment just above those lines.

I'm breaking things. Sorry.

I think I understood it now: the #ifndef protects only the check for TSC availability on non-NUMA archs. If it's available, and not under NUMA (so the ifndef), use it (jump to the rdtscll()), otherwise return the expression result.

The strange thing to me is that I'm getting 1/10 of the expected value when measuring tasks timeslices. Instead of getting ~100ms for tasks which just burn CPU, I'm getting 10ms.

I measure timeslices inside schedule(), updating the average timeslice for the leaving process, using (now - prev->timestamp).

Any clue of what am I doing wrong?

Regards.

unsigned long long sched_clock(void)

{
        unsigned long long this_offset;

        /*
         * In the NUMA case we dont use the TSC as they are not
         * synchronized across all CPUs.
         */
#ifndef CONFIG_NUMA
        if (!use_tsc)
#endif
                return (unsigned long long)jiffies * (1000000000 / HZ);

        /* Read the Time Stamp Counter */
        rdtscll(this_offset);

        /* return the value in ns */
        return cycles_2_ns(this_offset);
}
Signature · 7 lines
Ricardo Nabinger Sanchez
GNU/Linux #140696 [http://counter.li.org]
Slackware Linux

  Warning: 
    Trespassers will be shot.
    Survivors will be shot again.
#10 lore raw
Quoted 18 lines · Ricardo Nabinger Sanchez

Quoting Tim Schmielau <[email protected]> Sent on Fri, 28 Nov 2003 17:34:49 +0100 (CET)

Quoted 7 lines
Quoted 3 lines

This patch fixes an #endif misplacement, which leads to dead code in sched_clock() in arch/i386/kernel/timers/timer_tsc.c, due to a return outside the ifdef/endif.

No, this is exactly what is intended: don't use the TSC on NUMA, use jiffies instead.
Look at the comment just above those lines.

I'm breaking things. Sorry.

I think I understood it now: the #ifndef protects only the check for TSC availability on non-NUMA archs. If it's available, and not under NUMA (so the ifndef), use it (jump to the rdtscll()), otherwise return the expression result.

Quoted 8 lines

The strange thing to me is that I'm getting 1/10 of the expected value when measuring tasks timeslices. Instead of getting ~100ms for tasks which just burn CPU, I'm getting 10ms.

I measure timeslices inside schedule(), updating the average timeslice for the leaving process, using (now - prev->timestamp).

Any clue of what am I doing wrong?

Hi,

It could it be possible that TSC isn't being enabled on your processor due to a buggy TSC. You can try to trace the code on init_tsc(), in arch/i386/kernel/timers/timer_tsc.c to realize what's going on.

You can also try to check whether you're using HPET or not (when enabled and supported it makes use of TSC) in your .config . In the case you're not using it, the code that calibrates and dictates how to deal with the stamp counter is calibrate_tsc(), in arch/i386/kernel/timers/common.c.

Hope this helps,
Lucas