Re: [PATCH jk/pkt-line-cleanup] t5700-clone-reference: send trace to fd 2, not 3, to please Windows git

From: Johannes Sixt <j.sixt@viscovery.net>
Date: Mar 20, 2013, 11:30 UTC
Message-ID: <51499DE7.2030404@viscovery.net>
In-Reply-To: <20130320093313.GA9482@sigill.intra.peff.net>

Am 3/20/2013 10:33, schrieb Jeff King:

Show 19 quoted lines

> On Wed, Mar 20, 2013 at 09:24:44AM +0100, Johannes Sixt wrote:
> 
>> From: Johannes Sixt <j6t@kdbg.org>
>>
>> Two tests use GIT_TRACE=3 to dump debugging information of git. On
>> Windows, however, bash is unable to set up file descriptor 3 correctly
>> for its child process, so that git reports "Bad file descriptor" on
>> every trace attempt. The 'git clone' test succeeds nevertheless because
>> an empty trace file remains, and there is only a check for the
>> absence of a particular line. But the 'git fetch' process ultimately
>> hangs (the dynamics that lead to this surprising result have not been
>> investigated).
>>
>> Since we do not otherwise use stderr in the test cases, divert the
>> trace dump to stderr.
> 
> I think that is OK, but I'm curious why this is a problem _now_, and not
> with the code prior to 97a83fa8. The old GIT_DEBUG_SEND_PACK was also
> just calling write() to descriptor 3.

Before this change, both affected commands completed and the trace file was empty. Notice that in both test cases we only check for the absence of certain lines, which is naturally true for an empty file, so that the tests pass.

With the updated code, 'git fetch' hung, which is how I noticed the problem. As I said, I didn't investigate where and why this happens.

>>  This fixes a regression introduced in 
> 
> ...in 97a83fa8, I assume.

Yes, correct. (Sorry, I got distracted while I looked up the commit.)

-- Hannes