Re: Bug: Failure if stdin is closed.

From: Thomas Rast <trast@inf.ethz.ch>
Date: Jul 11, 2013, 15:39 UTC
Message-ID: <87r4f52jjf.fsf@linux-k42r.v.cablecom.net>
In-Reply-To: <201307111511.r6BFBO44010988@freeze.ariadne.com>

worley@alum.mit.edu (Dale R. Worley) writes:

Show 8 quoted lines

> (The original problem and the discussion that ensued is on the
> git-users mailing list:
> https://groups.google.com/forum/#!topic/git-users/lNQ7Cn35EqA)
>
> "git commit" (and probably other operations) fail if standard input
> (fd 0) is closed when git starts.  A simple test case follows.  (The
> execution is version 1.7.7.6, but the code listed below is from a very
> recent commit.)

While you found a real bug in the use of open() below, I'm not convinced that this scenario is something that should be guarded against. For example [1] says

  At program start-up, three streams shall be predefined and need not be
  opened explicitly: standard input (for reading conventional input),
  standard output (for writing conventional output), and standard error
  (for writing diagnostic output).

I haven't found a similar clause relating to the file descriptors 0--2, but my reading of the above paragraph is that running a program without an open stdin, stdout and stderr is, in fact, not POSIX.

Closing 2 usually has even funkier consequences if the program proceeds to open some other file, it ends up as fd 2, and it then dies with an error. In that sense it might be saner to simply die whenever open() gives an FD in the 0..2 range (and we weren't explicitly trying to reopen one of them).

May I ask why you need this, and to what extent this problem cannot be solved by instead redirecting from/to /dev/null?

> The fundamental error is that git_mkstemps_mode() in wrapper.c assumes
> that if open() is successful, its return value must be >0.  That is
> not true, because if fd 0 is closed, then open() can successfully
> return 0.  I have not tested it, but it looks like the fix is:

[...]

Show 8 quoted lines

> 		 template[5] = letters[v % num_letters]; v /= num_letters;
>
> 		 fd = open(pattern, O_CREAT | O_EXCL | O_RDWR, mode);
> -		 if (fd > 0)
> +		 if (fd >= 0)
> 			 return fd;
> 		 /*
> 		  * Fatal error (EPERM, ENOSPC etc).

This is an actual bug, as you say; open() is defined to return -1 on error, not 0. You can turn this into a patch, and send it as per Documentation/SubmittingPatches.

However, does it fully fix the issue you describe? What if you then run 'git checkout -F -' to read the message from stdin?

Show 5 quoted lines

> However, when looking at the code, I noticed that few of the functions
> have comments describing what they do, and none describe their input
> and output values.  In particular, there are no comments specifying
> what the error return values are.  This is appalling for a supposedly
> professional-quality project!

You are touching on a sore point of the git code base. Some contributors have made a point of adding comments where appropriate, so we're improving, but round tuits are in short supply as always. If you can supply such tuits, they would be appreciated.

[1]  http://pubs.opengroup.org/onlinepubs/009695399/functions/stdin.html

-- 
Thomas Rast
trast@{inf,student}.ethz.ch