Re: [PATCH] Fix start_command closing cmd->out/in regardless of cmd->close_out/in

From: Junio C Hamano <gitster@pobox.com>
Date: Nov 21, 2007, 09:11 UTC
Message-ID: <7vmyt8gdzv.fsf@gitster.siamese.dyndns.org>
In-Reply-To: <46dff0320711201838g5affba6bo21a8c837b0bef681@mail.gmail.com>

"Ping Yin" <pkufranky@gmail.com> writes:

Show 5 quoted lines

> On Nov 21, 2007 12:17 AM, Johannes Sixt <j.sixt@viscovery.net> wrote:
> ...
>> This way the change is more local without affecting well-tested other callers.
>
> This way works, but it is a tricky one, not a natural or graceful one.

I do not know about "natural". That largely would depend on where one starts thinking about the issues from.

But I think an API definition that says "These fds are closed after the call, so if you are going to use them, you can dup() them beforehand" is equally valid, and I suspect that forgetting to dup() is easier to detect than forgetting to close() --- you will notice the former mistake immediately because your read and write say "oops, nobody on the other end" but the latter mistake will result in a hung process. And for that reason, I think it can be called more "graceful". So ...

Show 9 quoted lines

>> Furthermore, I don't think that it's correct to just set the .close_in or
>> .close_out flags. This will close the fd only in finish_command(), which can
>> be too late: Think again of a writable pipe end that remains open and keeps
>> the reader waiting for input that is not going to happen.
>
> This may happen. However, i have scanned all the git codes using the
> auto closing behaviour and i don't discover the problem you mentioned.
> So i think it deserves to correct the misbehaviour after carefully
> testing. And we can make a clarification for that if necessary.

... I do not necessarily agree that your patch is correcting the misbehaviour.