[PATCH 1 of 6] dirstate: make dirstate.write() callers pass repo object to it

[Matt Mackall]

On Thu, 2015-10-15 at 02:35 +0900, FUJIWARA Katsunori wrote:
> # HG changeset patch
> # User FUJIWARA Katsunori <foozy at lares.dti.ne.jp>
> # Date 1444843899 -32400
> #      Thu Oct 15 02:31:39 2015 +0900
> # Node ID 59a9e52b17e1af3f69b6c9472c9c10a6b00d9b24
> # Parent  79d86ab65c9def3fdd65ec972bc5fa89688a19ff
> dirstate: make dirstate.write() callers pass repo object to it
> 
> Now, 'dirstate.write(repo)' delays writing in-memory changes out, if
> any transaction is running.

I guess I haven't been paying close enough attention to this as I didn't
notice the patch that actually introduced this parameter.

We should never pass a repo object to dirstate. It's a layering
violation that we've carefully avoided for 10 years and makes it way too
easy to get dirstate.py more entangled with upper levels that it has no
business knowing about.

Passing a transaction object by itself is ok, as it's a much lower level
object that dirstate has perfectly good reasons to know about.

Similar concerns exist for passing a repo object to a ui, another
situation we've been carefully avoiding since the beginning of time.
We definitely do not want ui objects poking at repo objects. They
probably shouldn't have code specific to pending transactions on general
principle either.

I'm going to put the currently-in-flight patches here on hold to give
you a chance to look at the alternatives.

FYI, there are some (old) docs on the layering rules:

https://www.mercurial-scm.org/wiki/WhatGoesWhere

Not by accident, this more or less agrees with the import structure:

https://www.selenic.com/blog/?p=626  (graph at the bottom)

-- 
Mathematics is the supreme nostalgia of our time.