Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1380658 > unrolled thread

Re: [PATCH v2] drm/msm: Use 64-bit timekeeping

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH v2] drm/msm: Use 64-bit timekeeping Arnd Bergmann <arnd@arndb.de> - 2016-04-17 01:50 +0200
    Re: [PATCH v2] drm/msm: Use 64-bit timekeeping Tina Ruchandani <ruchandani.tina@gmail.com> - 2016-04-21 13:40 +0200
      Re: [PATCH v2] drm/msm: Use 64-bit timekeeping Arnd Bergmann <arnd@arndb.de> - 2016-04-21 14:30 +0200
    Re: [PATCH v2] drm/msm: Use 64-bit timekeeping Tina Ruchandani <ruchandani.tina@gmail.com> - 2016-04-21 13:40 +0200

#1380658 — Re: [PATCH v2] drm/msm: Use 64-bit timekeeping

On Wednesday 13 April 2016 02:52:14 Tina Ruchandani wrote:
>                         ktime_t rem = ktime_sub(*timeout, now);
> -                       struct timespec ts = ktime_to_timespec(rem);
> -                       remaining_jiffies = timespec_to_jiffies(&ts);
> +                       struct timespec64 ts = ktime_to_timespec64(rem);
> +
> +                       remaining_jiffies = timespec64_to_jiffies(&ts);
> 

Hi Tina,

The change looks correct to me, but I wonder if we should optimize
this code a little more, as it does two expensive 64-bit divisions.

How about

	remaining_jiffies = msecs_to_jiffies(ktime_ms_delta(*timeout, now));

which only does one 64-bit division, and it's one that we can probably
optimize out in the future (we can check in ktime_ms_delta whether the
difference is more than 2^32 nanoseconds as the fast path).

	Arnd

[toc] | [next] | [standalone]

#1384106

>> which only does one 64-bit division, and it's one that we can probably
>> optimize out in the future (we can check in ktime_ms_delta whether the
>> difference is more than 2^32 nanoseconds as the fast path).

It looks like ktime_divns already has that optimization for 32-bit divisor,
so your solution should avoid the 64-bit division.

[toc] | [prev] | [next] | [standalone]

#1384149

On Thursday 21 April 2016 04:39:04 Tina Ruchandani wrote:
> >> which only does one 64-bit division, and it's one that we can probably
> >> optimize out in the future (we can check in ktime_ms_delta whether the
> >> difference is more than 2^32 nanoseconds as the fast path).
> 
> It looks like ktime_divns already has that optimization for 32-bit divisor,
> so your solution should avoid the 64-bit division.

I meant an optimization for a 32-bit dividend, not divisor,
e.g. doing:

diff --git a/include/linux/ktime.h b/include/linux/ktime.h
index 2b6a204bd8d4..4fbf735ec0af 100644
--- a/include/linux/ktime.h
+++ b/include/linux/ktime.h
@@ -169,13 +169,17 @@ static inline bool ktime_before(const ktime_t cmp1, const ktime_t cmp2)
 extern s64 __ktime_divns(const ktime_t kt, s64 div);
 static inline s64 ktime_divns(const ktime_t kt, s64 div)
 {
+	s64 ns = kt.tv64;
+
 	/*
 	 * Negative divisors could cause an inf loop,
 	 * so bug out here.
 	 */
 	BUG_ON(div < 0);
-	if (__builtin_constant_p(div) && !(div >> 32)) {
-		s64 ns = kt.tv64;
+
+	if ((ns >> 32) == 0) {
+		return (s32)ns / div;
+	else if (__builtin_constant_p(div) && !(div >> 32)) {
 		u64 tmp = ns < 0 ? -ns : ns;
 
 		do_div(tmp, div);

I also just looked at the implementation of do_div() in
include/asm-generic/div64.h, and it already does that for
non-constant divisors, but I don't understand __div64_const32()
enough to know if the compiler end up doing the same
optimization for the constant divisor we have here.

	Arnd

[toc] | [prev] | [next] | [standalone]

#1384108

>
> How about
>
>         remaining_jiffies = msecs_to_jiffies(ktime_ms_delta(*timeout, now));
>
> which only does one 64-bit division, and it's one that we can probably
> optimize out in the future (we can check in ktime_ms_delta whether the
> difference is more than 2^32 nanoseconds as the fast path).

Hi Arnd,
I had thought about that, but discard that approach being confused
about the truncation.
ktime_ms_delta returns s64 and msecs_to_jiffies will truncate that
input to int. However,
I now realize that for the msecs value to be greater than 32 bits, the
time delta has to be
>= ((2^29)/(60*60*24*365)) or >= 17 years. So your solution is safe.
If that sounds ok, I will send out a v3.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web