Re: psql ignores failure to open -o target file

I wrote:
> I just noticed that parse_psql_options() ignores the result of setQFout(),
> meaning that if the argument of a -o command line option is bogus, we'll
> ignore the switch entirely after printing an error report.

There's more silliness in the same area. \o with an invalid target spec
is treated like \o with no argument, ie, revert the target to stdout:

regression=# \o somefile
regression=# select 1; -- output goes to somefile, as expected
regression=# \o /dev/bogus
/dev/bogus: Permission denied
regression=# select 1;
?column?
----------
1
(1 row)

Surely that should have no effect beyond printing the error message.

Also, I notice that we try to set SIGPIPE handling to SIG_IGN whenever
the \o target is a pipe, which is good; but places that transiently
change SIGPIPE handling will unconditionally reset it back to SIG_DFL.
One such place is ClosePager(). That bug seems only latent because we
only invoke a pager when queryFout == stdout, so we would never get
there while the target is a pipe. But another such place is do_copy(),
and we definitely can call that while \o is set to a pipe.

In short, a bit of refactoring is needed around setQFout...

regards, tom lane