Re: [PATCH 0/4] git --paginate: do not commit pager choice too early

From: Jeff King <peff@peff.net>
Date: Jun 28, 2010, 10:22 UTC
Message-ID: <20100628102201.GA24817@coredump.intra.peff.net>
In-Reply-To: <20100628101335.GA5007@burratino>

On Mon, Jun 28, 2010 at 05:13:35AM -0500, Jonathan Nieder wrote:

Show 11 quoted lines

> > But reading the message for patch 4/4, I can't help but wonder if the
> > right way forward is for the git wrapper to _always_ find the repository
> > as the very first thing.
> [...]
> > the worst case should be a little bit of
> > wasted effort.
> 
> This is a very useful thought.  Git without the chdir() to toplevel
> would be pleasanter, I think.  (Yes, I realize you weren’t necessarily
> suggesting suppressing a chdir_to_toplevel() in place of
> setup_git_repository() and friends.)

I had just meant that we would not do the chdir() initially, but would do so right before running the actual command which wanted repository setup (and commands like init which do not do that setup would never chdir to the toplevel). But we must always run at least aliases and external sub-commands from the toplevel to keep backwards compatibility.

Show 16 quoted lines

> Regarding the repository search: automounters can cause pain if
> DISCOVERY_ACROSS_FILESYSTEM is set and GIT_CEILING_DIRECTORIES is not
> set appropriately.  An important script that does
> 
>   tmpdir=$(mktemp -d)
>   cd "$tmpdir"
>   ...
> 
>   git diff --no-index a b
>   ...
> 
> could hang indefinitely if some nut has created a named pipe at
> /tmp/.git.  I haven’t come up with any non far-fetched reason to mind
> the repository search, though.
> 
> Will think about this.

Yeah, I considered that, too. But I doubt it is a big problem. If it were, then just running "git log" when you are not in a repository would be similarly painful. The people who have setups like that are already having to deal with GIT_CEILING_DIRECTORIES to make things less painful (though I imagine the new cross-filesystem discovery code will prevent most of the pain without them having to actually do anything).

So yes, we are adding the extra lookup for commands like "git clone", but I suspect in practice nobody will care. If it is a big deal, we can do something like:

  if (!strcmp(cmd, "clone") || !strcmp(cmd, "init"))
     ... don't do setup ...

which I admit is a terrible hack, but it is an optimization, not a correctness thing. So we are not shutting out external user-defined commands that might not care about the repo. We are just making our common built-in ones a bit more efficient.

-Peff