Re: [PATCH 0/6] Use time_t

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Feb 28, 2017, 23:10 UTC
Message-ID: <alpine.DEB.2.20.1703010000260.3767@virtualbox>
In-Reply-To: <6bd6ef64-2749-f17a-aea8-d25b147892c5@web.de>

Hi René,

On Tue, 28 Feb 2017, René Scharfe wrote:

Show 41 quoted lines

> Am 28.02.2017 um 21:54 schrieb Johannes Schindelin:
> >
> > On Tue, 28 Feb 2017, Junio C Hamano wrote:
> >
> > > René Scharfe <l.s.r@web.de> writes:
> > >
> > > > Am 28.02.2017 um 15:28 schrieb Jeff King:
> > > >
> > > > > It looks from the discussion like the sanest path forward is our
> > > > > own signed-64bit timestamp_t. That's unfortunate compared to
> > > > > using the standard time_t, but hopefully it would reduce the
> > > > > number of knobs (like TIME_T_IS_INT64) in the long run.
> > > >
> > > > Glibc will get a way to enable 64-bit time_t on 32-bit platforms
> > > > eventually
> > > > (https://sourceware.org/glibc/wiki/Y2038ProofnessDesign).  Can
> > > > platforms that won't provide a 64-bit time_t by 2038 be actually
> > > > used at that point?  How would we get time information on them?
> > > > How would a custom timestamp_t help us?
> > >
> > > That's a sensible "wait, let's step back a bit".  I take it that you
> > > are saying "time_t is just fine", and I am inclined to agree.
> > >
> > > Right now, they may be able to have future timestamps ranging to
> > > year 2100 and switching to time_t would limit their ability to
> > > express future time to 2038 but they would be able to express
> > > timestamp in the past to cover most of 20th century.  Given that
> > > these 32-bit time_t software platforms will die off before year 2038
> > > (either by underlying hardware getting obsolete, or software updated
> > > to handle 64-bit time_t), the (temporary) loss of 2038-2100 range
> > > would not be too big a deal to warrant additional complexity.
> >
> > You seem to assume that time_t is required to be signed. But from my
> > understanding that is only guaranteed by POSIX, not by ISO C.
> >
> > We may very well buy ourselves a ton of trouble if we decide to switch
> > to `time_t` rather than to `int64_t`.
> 
> True, and time_t doesn't even have to be an integer type.  But which
> platforms capable of running git use something else than int32_t or
> int64_t?

That kind of thinking is dangerous. We don't know what platforms are running Git, and we have a very clear example how we got it very wrong recently, when we broke building with musl by requiring REG_STARTEND support [*1*].

So why gamble? If we switch to uint64_t, it would definitely provide the smoothest upgrade path. It is what the code assumed implicitly when we broke 32-bit in v2.9.1.

If anybody really wants to support negative timestamps, it should be done on top of my work. My current patch series does not even start to try to address the ramifications of negative timestamps (see e.g. the use of strtoull() for parsing). It is quite unreasonable to ask for such a fundamental design change when it could very easily be done incrementally instead, when needed, by someone who needs it.

My work would pave the way for that effort, of course. But this is really as far as I can go with this patch series, given that I have bigger fish to fry than to support negative timestamps.

Ciao, Dscho

Footnote *1*: I still deeply regret deviating from my v1 that did *not* require REG_STARTEND, but would have kept things working for platforms without REG_STARTEND by simulating it.

But our thinking was: who would want to run Git in an environment so ridiculously old that it does not have that clearly useful REG_STARTEND support? Our answer was "nobody". And it was incorrect.