2016-04-13 12:09:08

by Tina Ruchandani

[permalink] [raw]
Subject: [PATCH v2] drm/msm: Use 64-bit timekeeping

'struct timespec' uses a 32-bit seconds which will overflow in year
2038 and beyond. This patch replaces timespec with timespec64. The
code is correct as is - the patch is merely part of a larger attempt
to remove all 32-bit timekeeping variables (timespec, timeval, time_t)
from the kernel.

Signed-off-by: Tina Ruchandani <[email protected]>
--
Changes in v2:
Fix checkpatch warning
---
drivers/gpu/drm/msm/msm_drv.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
index c03b967..59c1948 100644
--- a/drivers/gpu/drm/msm/msm_drv.c
+++ b/drivers/gpu/drm/msm/msm_drv.c
@@ -717,8 +717,9 @@ int msm_wait_fence(struct drm_device *dev, uint32_t fence,
remaining_jiffies = 0;
} else {
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);
}

if (interruptible)
--
2.8.0.rc3.226.g39d4020


2016-04-16 23:47:41

by Arnd Bergmann

[permalink] [raw]
Subject: 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

2016-04-21 11:35:52

by Tina Ruchandani

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

>
> 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.

2016-04-21 11:39:08

by Tina Ruchandani

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

>> 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.

2016-04-21 12:21:45

by Arnd Bergmann

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

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