Re: [PATCH 1/2] merge: close the index lock when not writing the new index

From: Martin Ågren <martin.agren@gmail.com>
Date: Nov 13, 2017, 21:12 UTC
Message-ID: <CAN0heSrFdGQ9469ugQwmjP5ieoxb5NCR3Kpz6ZQ4J5TyarwQ6w@mail.gmail.com>
In-Reply-To: <20171110231314.30711-2-joel@teichroeb.net>

On 11 November 2017 at 00:13, Joel Teichroeb <joel@teichroeb.net> wrote:

> If the merge does not have anything to do, it does not unlock the index,
> causing any further index operations to fail. Thus, always unlock the index
> regardless of outcome.

>         if (clean < 0)
>                 return clean;

Do we need to roll back the lock also if `clean` is negative? The current callers are built-ins which will error out, but future callers might be caught off guard by this.

Show 12 quoted lines

> -       if (active_cache_changed &&
> -           write_locked_index(&the_index, &lock, COMMIT_LOCK))
> -               return err(o, _("Unable to write index."));
> +       if (active_cache_changed) {
> +               if (write_locked_index(&the_index, &lock, COMMIT_LOCK))
> +                       return err(o, _("Unable to write index."));
> +       } else {
> +               rollback_lock_file(&lock);
> +       }
>
>         return clean ? 0 : 1;
>  }

Looks correct. A simpler change which would still match the commit message would be to unconditionally call `rollback_lock_file()` just before returning. That would perhaps be slightly more future-proof, since it will always leave the lock unlocked, even if the if-else grows more complicated.

Well, "always" modulo returning early and forgetting to roll back the lock. ;-) Looking at existing code, it's not obvious which way we should prefer. Just a thought.

Thanks for spotting this. I was poking around here recently, but failed to notice this lax lock-handling.

Martin