Web lists-archives.com

Re: Summary of the 2038 BoF at DC17




On Mon, 2017-09-18 at 18:33:12 +0100, Steve McIntyre wrote:
> On Mon, Sep 18, 2017 at 05:46:34PM +0100, Ian Jackson wrote:
> >Steve McIntyre writes ("Re: Summary of the 2038 BoF at DC17"):
> >> It depends on how/where/why you're embedding 64-bit time,
> >> basically. If you're embedding a time_t (or a struct including a
> >> time_t) in your ABI and want to keep to something similar, it's worth
> >> waiting to see what's going to be standardised then using that.
> >
> >Are you saying that if I am designing an API/ABI now I should write:
> >
> >  typedef struct {
> >    blah blah;
> >    time_t whenever;
> >    blah blah;
> >  } MyAPIThing;
> >
> >rather than
> >
> >  typedef struct {
> >    blah blah;
> >    uint64_t whenever;
> >    blah blah;
> >  } MyAPIThing;
> >
> >?  Really ?
> >
> >I think that's bad advice.

I have to agree here. I was also suprised by that recommendation in the
initial report.

> Yes, really. You've now hidden that you're storing time data by using
> another data type, which makes things much harder to find if anybody
> else is scanning for time-handling code. And you've made assumptions
> about how new time-handling APIs are likely to look in the near-ish
> future when people have worked everything out and agreed new
> standards. If the new stuff ends up using a different representation
> with 96 or even 128 bits in total, I'd argue that it's cleaner to wait
> for that and not gamble.

While using more semantic types would be nicer, and more searchable,
the reality is that we've got already systems out there with 64-bit
time_t/clock_t (OpenBSD for example). If POSIX (say) standardized on a
new time96_t or similar, code using 64-bit time would still be no worse
off, it would actually still be in a better position compared to 32-bit
time.

In addition, for the same reason that exposing off_t directly as part
of an API is a very bad idea, because you are then requiring the users
to match your build LFS state, or they will break horribly. Doing the
equivalent with time_t, assuming there cannot be a scorched-earth
just-rebuild-the-world approach to transitioning to non 32-bit time,
will be also a bad idea. (We still do not have 100% LFS coverage!)

The alternative is to provide interfaces similar to glibc, for example
one for off_t (32-bit) another for off64_t, etc, which is also horrible.

> >I would do the latter.  Even though that means writing library code
> >internally that checks whether the supplied value of `whenever' fits
> >in whatever the system calls a time_t.
> 
> Your code, your choice...

Well for public APIs, it's not just "your code", it's code that
affects all its users.

Thanks,
Guillem