Re: [PATCH v3 02/25] Convert git_snpath() to strbuf_git_path()

From: Junio C Hamano <gitster@pobox.com>
Date: Feb 20, 2014, 18:54 UTC
Message-ID: <xmqq1tyx7gxx.fsf@gitster.dls.corp.google.com>
In-Reply-To: <CACsJy8DB+OP6_vwZKWXjq4YbtiKDg2NE2zRsAuz1ER6oTmyGag@mail.gmail.com>

Duy Nguyen <pclouds@gmail.com> writes:

> OK so your question was if there was a git_path() or mkpath() call
> earlier in update_refs_for_switch() and the result was expected to
> remain stable till the end of update_refs_for_switch(), then this
> conversion could ruin it, correct? I didn't think about that,...

Yeah, I couldn't tell if you thought about it, and that was why I asked.

If a (recursively) caller does this:

	caller () {
		const char *path1 = git_path(...);
                const char *path2 = mkpath(...);
                const char *path3 = git_path_submodule(...);
		callee();
                use(path1, path2, path3);
	}

it was safe back when the callee() did not mess with the round-robin pathname buffer, but it will be broken once callee() does. While looking at the patch I didn't check what the caller was doing hence my question.

In general, in order to reduce that kind of hard-to-debug bugs, we should be reducing the uses of these functions when we do not have to (which applies equally to such a caller that expects multiple temporary paths to persist, and to a callee as well). Adding multiple repeated calls to git_path(), especially two of them formatting the same string into two separate round-robin pathname buffer, looked strange in a patch that was supposed to be a preparatory code-cleanup stage.