On Wed, Mar 29, 2017 at 10:06:52AM -0700, Junio C Hamano wrote:
> > This shows that we should be careful not to use git_path() in
> > freshen_shared_index(). It is using a shared buffer that can
> > too easily lead to races.
>
> The impression I get from the symptom is that after git_path() is
> called here, before check_and_freshen_file() uses that result, it
> (or functions it calls) uses git_path(), and the number of times it
> does so has changed since cc/split-index-config was written on the
> mainline, and the rotating 4-element buffer get_pathname() gives is
> now exhausted, leading to the failure you observed. By the way,
> that does not sound a race to me.
>
> In any case, that explains why bisect says the merge is the first
> bad one, and cures the confused reader ;-) The use of git_path() on
> the topic was still safe; it was a timebomb waiting to go off. The
> mainline started using more calls and the merge result was unsafe.
Yeah, it looks like that is what happened. I see that Christian bisected
the rebase to find the commit in the series that introduces the problem.
I'm mildly curious which commit upstream created the problem[1].
There's a reasonable chance it's some innocent-looking cleanup (possibly
one of my recent "stop using a fixed buffer" ones).
But in the end it doesn't really matter. I think code like:
const char *filename = git_path(...);
or
nontrivial_function(git_path(...));
is an anti-pattern. It _might_ be safe, but it's really hard to tell
without following the complete lifetime of the return value. I've been
tempted to suggest we should abolish git_path() entirely. But it's so
darn useful for things like unlink(git_path(...)), or other direct
system calls.
As an aside, this kind of static-buffer reuse _used_ to mean you might
see somebody else's buffer. Which is bad enough. But since the move to
use strbufs underneath the hood of git_path(), it may produce that
effect or it may be a use-after-free (if the strbuf had to reallocate to
grow in the meantime).
Anyway. The fix in the patch is obviously the right thing.
-Peff
[1] I think we could pinpoint the upstream change that caused the bad
interaction by bisecting between the merge-base and the first-parent
of the broken merge. For each commit, cherry-pick the complete
series on top of it, and test the result.