Re: [PATCH] vdso: Define NSEC_PER_SEC as 64-bit to prevent overflow

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, Aug 09 2025 at 17:49, Wake Liu wrote:

Please do not top post and trim your replies.

> CLOCK_REALTIME         ABSTIME PERIODIC timer fired early:       1 : [FAILED]
> Looks like there is some integer overflow on 32 bit builds in the
> timespec_sub function:
>
> long long timespec_sub(struct timespec a, struct timespec b)

That's in selftests and not in the kernel VDSO code.

Can you please be precise with your subject and changelog? You claimed:

 "This can lead to incorrect time calculations within the VDSO"

No, it can't not because the VDSO does not use it that way.

> {
> long long ret = NSEC_PER_SEC * b.tv_sec + b.tv_nsec;
>
> ret -= NSEC_PER_SEC * a.tv_sec + a.tv_nsec;
> return ret;
> }
>
> on 32 bit builds NSEC_PER_SEC and b.tv_sec are only 32 bit values
> which I'm guessing is causing the overflow.

Guessing is not a valid technical problem solving approach.

> NSEC_PER_SEC is defined as 1000000000L but if we change this to
> 1000000000LL then the test starts passing.

Yes, that function bogus on 32bit.

There are at least ten different ways to implement this function 32-bit
safe, but you picked the worst option of all, which breaks the kernel
build as you got told by 0-day.

I'm sure you can figure out how to ensure that the multiplication is
actually expanding to 64-bit or use some other approach to calculate
that delta.

And please make that an inline function and put it into a header file in
the timer selftest directory, and replace _all_ copies of it in the
various timer tests.

Thanks,

        tglx