Re: Bug: git branch --unset-upstream command can nuke config when disk is full.

From: Jeff King <peff@peff.net>
Date: Sep 13, 2017, 15:22 UTC
Message-ID: <20170913152205.sw3i5xzsynx62nea@sigill.intra.peff.net>
In-Reply-To: <CANgJU+XvJA2LYgFgmJZw7BR4-7qobvOh5zOSAVzGFOTsJ=Z8ug@mail.gmail.com>

On Wed, Sep 13, 2017 at 05:18:56PM +0200, demerphq wrote:

Show 15 quoted lines

> > Hmph. That is very disturbing. But with that information I should be
> > able to track down the culprit. Thanks for digging.
> 
> FWIW, I see that git_config_set_multivar_in_file_gently() uses
> write_in_full() which in turn uses xwrite(), but the latter has the
> following comment on it:
> 
> /*
>  * xwrite() is the same a write(), but it automatically restarts write()
>  * operations with a recoverable error (EAGAIN and EINTR). xwrite() DOES NOT
>  * GUARANTEE that "len" bytes is written even if the operation is successful.
>  */
> 
> I suspect that at this point I am not adding much value here, so I
> will leave it at this.

No, the problem is in this line:

                                 if (write_in_full(fd, contents + copy_begin,
                                                   copy_end - copy_begin) <
                                     copy_end - copy_begin)
                                          goto write_err_out;

write_in_full() returns -1 on error (_not_ how many bytes were actually written). So its return is a signed ssize_t. But the result of the pointer subtraction "copy_end - copy_begin" is an unsigned ptrdiff_t. The compiler promotes the signed to an unsigned, so the condition can never be true (the "-1" becomes the highest possible value).

I have the fix, but I'm searching the code base for other instances of the same error.

-Peff