Re: COPY FROM STDIN behaviour on end-of-file

Thomas Munro <thomas(dot)munro(at)enterprisedb(dot)com> writes:
> On Tue, May 16, 2017 at 6:29 PM, Vaishnavi Prabakaran
> <vaishnaviprabakaran(at)gmail(dot)com> wrote:
>> Hi, I could not reproduce this issue. Even after Ctrl+d , subsequent COPY
>> from commands reads the input properly. Is there any specific step you
>> followed or can you share the sample testcase?

> Hmm. Doesn't happen on GNU/Linux, does happen on macOS and FreeBSD.

Hah, confirmed here. Thinking about it, it seems like glibc must be
exceeding its authority to make this work on Linux; I don't know of
anything in POSIX saying that the eof indicator should go away without
a clearerr() call. In fact, it seems like glibc is violating its own
documentation --- "man feof" saith

The function feof() tests the end-of-file indicator for the stream
pointed to by stream, returning non-zero if it is set. The end-of-file
indicator can only be cleared by the function clearerr().

I had been supposing that this was a feature addition and should be left
for the next commitfest. But given that it already works as-expected on
popular platform(s), the fact that it doesn't work the same on some other
platforms seems like a portability bug rather than a missing feature.
Now I'm inclined to treat it as a bug and back-patch.

Reviewing the actual patch, though ... there seem to be paths through
handleCopyIn that would not hit this, particularly the sigsetjmp path.
It's unlikely that we'd get control-C before we finish servicing
control-D, but maybe not impossible. Wouldn't it be better just to do
an unconditional clear at the end, maybe

copyin_cleanup:

+ /* In case input is from a tty, reset the EOF indicator. */
+ clearerr(copystream);
+
/*
* Check command status and return to normal libpq state.
*

(I'd be inclined to make the comment significantly more verbose than
that, btw.)

regards, tom lane