Re: [PATCH 6/6] vfs: change inode times to use struct timespec64

[Kees Cook]

On Mon, May 14, 2018 at 10:25 AM, Deepa Dinamani <deepa.kernel@xxxxxxxxx> wrote:
> On Mon, May 14, 2018 at 9:30 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>> On Sun, May 13, 2018 at 9:05 PM, Deepa Dinamani <deepa.kernel@xxxxxxxxx> wrote:
>>> Kees mentioned that he wants to merge a patch to pstore that changes
>>> it to use timespec64 internally for 4.17:
>>> https://lkml.org/lkml/2018/5/13/3
>>
>> I'm still working on a v2 for pstore. What is the correct
>> cross-architecture format string for timespec64's tv_sec? In your
>> other patches, you're using %lld and a (long long) cast. I'd really
>> like to avoid the need for casts.
>
> We cannot really avoid it for now.
> struct timespec64 is defined this way for now:
>
> struct timespec {
> __kernel_time_t tv_sec; /* seconds */
> long tv_nsec; /* nanoseconds */
> };
>
> #if __BITS_PER_LONG == 64
> /* this trick allows us to optimize out timespec64_to_timespec */
> # define timespec64 timespec
>
> #else
>
> struct timespec64 {
> time64_t tv_sec; /* seconds */
> long tv_nsec; /* nanoseconds */
> };
>
> #endif
>
> This will all lead to tv_sec being long on a 64 bit architecture and
> long long on a 32 bit architecture.
> So there is no way of avoiding the cast for now.
>
> We plan to get rid of this trick and to have a single definition for
> timespec64. But, that cleanup is planned for later when we cleanup all
> struct timespec uses internally.

Can we do something like:

#if __BITS_PER_LONG == 64
# define TVSEC_FMT "%ld"
#else
# define TVSEC_FMT "%lld"
#endif

so we can do stuff like:  sprintf(buf, "seconds: " KTIME_FMT, time->tv_sec)

? It seems easier to clean up than casts.

-Kees

-- 
Kees Cook
Pixel Security