Re: Fix overflow in DecodeInterval

Joseph Koshakow <koshy44@gmail.com>

From: Joseph Koshakow <koshy44@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Andres Freund <andres@anarazel.de>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2022-04-03T15:23:07Z
Lists: pgsql-hackers
On Sun, Apr 3, 2022 at 3:09 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> I wrote:
> > Cool.  I've pushed the patch.
>
> Hmm ... buildfarm's not entirely happy [1][2][3]:
>
> diff -U3 /home/nm/farm/gcc64/HEAD/pgsql.build/src/test/regress/expected/interval.out /home/nm/farm/gcc64/HEAD/pgsql.build/src/test/regress/results/interval.out
> --- /home/nm/farm/gcc64/HEAD/pgsql.build/src/test/regress/expected/interval.out 2022-04-03 04:56:32.000000000 +0000
> +++ /home/nm/farm/gcc64/HEAD/pgsql.build/src/test/regress/results/interval.out  2022-04-03 05:23:00.000000000 +0000
> @@ -1465,7 +1465,7 @@
>   LINE 1: select interval 'PT2562047788.1:00:54.775807';
>                           ^
>   select interval 'PT2562047788:01.:54.775807';
> - ERROR:  interval field value out of range: "PT2562047788:01.:54.775807"
> + ERROR:  invalid input syntax for type interval: "PT2562047788:01.:54.775807"
>   LINE 1: select interval 'PT2562047788:01.:54.775807';
>                           ^
>   -- overflowing with fractional fields - SQL standard format
>
> What do you make of that?  I'm betting that strtod() works a
> bit differently on those old platforms, but too tired to
> look closer tonight.
>
>                         regards, tom lane
>
> [1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=sungazer&dt=2022-04-03%2004%3A56%3A34
> [2] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=hoverfly&dt=2022-04-03%2000%3A51%3A50
> [3] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=anole&dt=2022-04-03%2000%3A32%3A10

I think I know that the issue is. It's with `ParseISO8601Number` and
the minutes field "1.".
Previously that function parsed the entire field into a single double,
so "1." would
be parsed into 1.0. Now we try to parse the integer and decimal parts
separately. So
we first parse "1" into 1 and then fail to "." into anything because
it's not a valid decimal.

What's interesting is that I believe this syntax, "1.", always would
have failed for
non-ISO8601 Interval. It was only previously valid with ISO8601 intervals.

- Joe Koshakow



Commits

  1. Fix portability issues in datetime parsing.

  2. Fix overflow hazards in interval input and output conversions.

  3. Add a couple more tests for interval input decoding.