Re: Crash on MSYS2 with GIT_WORK_TREE

From: Junio C Hamano <gitster@pobox.com>
Date: Mar 8, 2017, 02:36 UTC
Message-ID: <xmqqh9344hq6.fsf@gitster.mtv.corp.google.com>
In-Reply-To: <alpine.DEB.2.20.1703080104580.3767@virtualbox>

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

Show 11 quoted lines

> The problem is actually even worse: On *Linux*, this happens:
>
> 	$ GIT_WORK_TREE=c:/invalid git rev-parse HEAD
> 	Segmentation fault (core dumped)
>
> The reason is this: when set_git_work_tree() was converted from using
> xstrdup(real_path()) to real_pathdup(), we completely missed the fact that
> the former passed die_on_error = 1 to strbuf_realpath(), while the latter
> passed die_on_error = 0. As a consequence, work_tree can be NULL now, and
> the current code does not expect set_git_work_tree() to return
> successfully after setting work_tree to NULL.

Ouch.

> Brandon, I have a hunch that pretty much all of the xstrdup(real_path())
> -> real_pathdup() sites have a problem now. The previous contract was that
> real_path() would die() if the passed path is invalid. The new contract is
> that real_pathdup() returns NULL in such a case.

OK, so it appears that we'd better audit all the callsites of real_pathdup() and see if anybody _assumes_ that the return values are not NULL. They all need fixing.

Thanks for digging it through to the root cause.