Skip to contentSkip to conversation

x86/tsc: improve arithmetic division

PATCH In review Wen Yang 2 participants 2 messages Jan 30, 2020 – Feb 1, 2020 x86

lore
Share & tools
lore

Wen Yang

author
#0 lore raw

PATCH 1/1 x86/tsc: improve arithmetic division

do_div() does a 64-by-32 division. Use div64_ul64() or div64_ul() instead of it if the divisor is 'ul64' or 'unsigned long', to avoid truncation to lower 32-bit.
And as a nice side effect also cleans up the function a bit.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
---
 arch/x86/kernel/tsc.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
arch/x86/kernel/tsc.c +2 -5
index 7e322e2daaf5..4c0320e68699 100644
@@ -357,9 +357,7 @@ static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
 	pm2 -= pm1;
 	tmp = pm2 * 1000000000LL;
 	do_div(tmp, PMTMR_TICKS_PER_SEC);
-	do_div(deltatsc, tmp);
-
-	return (unsigned long) deltatsc;
+	return (unsigned long) div64_u64(deltatsc, tmp);
 }
 
 #define CAL_MS		10
@@ -778,8 +776,7 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)
 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 
 		/* Check the reference deviation */
-		delta = ((u64) tsc_pit_min) * 100;
-		do_div(delta, tsc_ref_min);
+		delta = div64_ul(((u64) tsc_pit_min) * 100, tsc_ref_min);
 
 		/*
 		 * If both calibration results are inside a 10% window
Signature · 1 line
2.23.0
#1 lore raw
Quoted 42 lines · Wen Yang

do_div() does a 64-by-32 division. Use div64_ul64() or div64_ul() instead of it if the divisor is 'ul64' or 'unsigned long', to avoid truncation to lower 32-bit.
And as a nice side effect also cleans up the function a bit.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc "H. Peter Anvin" <[email protected]>

---
arch/x86/kernel/tsc.c | 7 ++-----

1 file changed, 2 insertions(+), 5 deletions(-)
arch/x86/kernel/tsc.c +1 -3
index 7e322e2daaf5..4c0320e68699 100644
@@ -357,9 +357,7 @@ static unsigned long calc_pmtimer_ref(u64 deltatsc,
 u64 pm1, u64 pm2)
 	pm2 -= pm1;
 	tmp = pm2 * 1000000000LL;
 	do_div(tmp, PMTMR_TICKS_PER_SEC);
-	do_div(deltatsc, tmp);
-
-	return (unsigned long) deltatsc;
+	return (unsigned long) div64_u64(deltatsc, tmp);
 }
 

#define CAL_MS 10

arch/x86/kernel/tsc.c +1 -2
@@ -778,8 +776,7 @@ static unsigned long
 pit_hpet_ptimer_calibrate_cpu(void)
 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 
 		/* Check the reference deviation */
-		delta = ((u64) tsc_pit_min) * 100;
-		do_div(delta, tsc_ref_min);
+		delta = div64_ul(((u64) tsc_pit_min) * 100, tsc_ref_min);
 
 		/*
Quoted 1 line

* If both calibration results are inside a 10% window

This is a *lot* more expensive on 32 bits (something like 10x) and as the output is truncated to unsigned long anyway, it is also unnecessary.

We don't use the remainder, so using do_div() is not merely unnecessary but almost certainly generates worse code: we are multiplying and then dividing by a constant, and most of the time gcc can optimize that into a single multiply/shift operation; otherwise we can do that optimization for it (see timeconst.bc.)

The one thing that gcc can't necessary do automatically is to know when a 64/32 → 32 division is safe; C semantics are truncation, but the CPU will trap. If it can turn it into a multiply then that problem obviously goes away.

So first I would test with regular / operators and see what code comes out.

Signature · 1 line
Sent from my Android device with K-9 Mail. Please excuse my brevity.