From: Jeff King <peff@peff.net>
Date: Mon, 02 Oct 2017 05:41:27 GMT
Subject: Re: [PATCH 05/11] cache-tree: simplify locking logic
Message-ID: <20171002054126.xkvv2zglefxcvbia@sigill.intra.peff.net>
In-Reply-To: <3475f0eee33ca0c4b2bc928191f0b06576ee29fb.1506862824.git.martin.agren@gmail.com>

On Sun, Oct 01, 2017 at 04:56:06PM +0200, Martin Ågren wrote:

> After we have taken the lock using `LOCK_DIE_ON_ERROR`, we know that
> `newfd` is non-negative. So when we check for exactly that property
> before calling `write_locked_index()`, the outcome is guaranteed.
> 
> If we write and commit successfully, we set `newfd = -1`, so that we can
> later avoid calling `rollback_lock_file` on an already-committed lock.
> But we might just as well unconditionally call `rollback_lock_file()` --
> it will be a no-op if we have already committed.
> 
> All in all, we use `newfd` as a bool and the only benefit we get from it
> is that we can avoid calling a no-op. Remove `newfd` so that we have one
> variable less to reason about.

Nice, this looks much simpler and the reasoning above is all sound.

I think cmd_checkout_index() has the exact same thing going on.

-Peff