From the department of punishing good deeds ... in commit 2dc4f011fd
I added setvbuf() calls to initdb to ensure that output to stdout and
stderr would appear in a consistent order regardless of whether the
output was going to a terminal or a file. The buildfarm shows that
on several (but not all) Windows machines, initdb is now failing
without printing anything. After a bit of research, I believe what
is happening is:
(1) On Windows, setvbuf interprets _IOLBF as _IOFBF.
(2) If _IOFBF is specified and size is zero, it reports EINVAL.
(3) If "parameter validation" is enabled, EINVAL turns into abort().
Thanks Microsoft for your careful attention to compliance with POSIX.
I see that syslogger.c encountered this problem long ago and solved
it by the expedient of using _IONBF not _IOLBF for the log output
file. That's probably what we must do for stdout in initdb as well.
It seems likely that we may need the same in other client programs
someday.
What I'm not totally sure about is how to wrap this up nicely.
I'm inclined first of all to move syslogger.c's LBF_MODE symbol to
port.h and rename it to, say, PG_IOLBF.
It might also be reasonable to create a wrapper macro along the line of
"PG_STD_IO_BUFFERING()" that would encapsulate the whole sequencesetvbuf(stdout, NULL, _IOLBF, 0);setvbuf(stderr, NULL,
_IONBF,0);
Or maybe we should have separate macros for those two calls. Or maybe
this is just a useless layer of abstraction and PG_IOLBF is enough
to make the calls portable.
Thoughts?
regards, tom lane