[PATCH 2 of 3] revert: rename working directory matcher 'm' -> 'wmatch'

Matt Harbison mharbison72 at gmail.com
Fri Mar 27 20:45:58 CDT 2015

On Fri, 27 Mar 2015 14:15:09 -0400, Matt Mackall <mpm at selenic.com> wrote:

> On Thu, 2015-03-26 at 21:06 -0400, Matt Harbison wrote:
>> On Thu, 26 Mar 2015 17:09:13 -0400, Matt Mackall <mpm at selenic.com>  
>> wrote:
>> > On Thu, 2015-03-26 at 00:20 -0400, Matt Harbison wrote:
>> >> # HG changeset patch
>> >> # User Matt Harbison <matt_harbison at yahoo.com>
>> >> # Date 1427333998 14400
>> >> #      Wed Mar 25 21:39:58 2015 -0400
>> >> # Node ID 781d27e7fcd3a438d02bf2e1a9ba0a85537da6b9
>> >> # Parent  6e1b26088b8eca13825a91e4e1e3547c2e89f6a0
>> >> revert: rename working directory matcher 'm' -> 'wmatch'
>> >>
>> >> Further down in the function, 'm' is reassigned to a matcher for a  
>> list
>> >> of
>> >> files.  At the bottom of the function, another matcher is created to
>> >> figure out
>> >> what subrepos need to be reverted.  For consistency with  
>> 5b85a5bc5bbb,
>> >> this will
>> >> be changed to a working directory matcher.
>> >
>> > Not sure about this one. The first use of 'm' here is the standard  
>> "make
>> > a matcher from the supplied args" pattern and the second definition
>> > appears to have very limited use.
>> >
>> > It seems the shadowing here is kind of intentional: we're making a  
>> more
>> > narrower version of the initial match that should be used for the rest
>> > of the function. So perhaps we should instead be moving the subrepo  
>> set
>> > calculation higher up instead of doing any renaming?
>> >
>> > But really, your patch 3 would actually be _simpler_ as a patch 2 that
>> > just did s/ctx/wctx/ locally, right? I'll just do that for now. Patch  
>> 1
>> > and 3' queued, thanks.
>> That looks like it works for this series.  The next thing to do though  
>> is
>> to pass a matcher to the subrepo, that is a narrowed form of the  
>> original
>> passed to the function.  The original matcher needs to make it to the
>> subrepo section, but moving the whole subrepo section to the top looks
>> like it would work too.
>> I guess now is a good time to ask before doing the work- subrepo revert
>> works by essentially doing a clean update of the subrepo, after doing a
>> cmdutil.revert() on that layer to generate backup files.  That
>> functionality would seem to preclude the path/into/subrepo form that is
>> being supported on more commands, and would seem to violate the  
>> documented
>> help for revert:
>> 	 "Because revert does not change the working directory parents..."
>> Granted, it isn't changing the parents of the outer repo, but the  
>> subrepo
>> working directory seems like an extension of the outer repo's working
>> directory.  Backout doesn't support subrepos, and maybe it isn't  
>> possible
>> to do so.  But it certainly can't if subrepos are being updated like  
>> this
>> by revert.  And because of the recursive nature, a subrepo N subrepos  
>> deep
>> will be clean updated N times.
> Oh man, that's a hard question. My internal model of subrepo semantics
> says that yes, we should get a clean checkout of subrepos under the top
> level, but I can believe that would surprise people. I'm not sure if the
> alternative is particularly logically consistent though.

I'm not sure I understand the logical consistency part.  If I want to  
revert a directory in a flat repo, that's spelled 'hg revert dir/subdir'.   
The result may or may not be consistent codewise- other code depending on  
subdir may now be missing a dependency, but that's what I asked for, and  
the tool doesn't get in the way.

Given that a subdirectory can be reverted (and because other commands  
support path/into/subrepo syntax), I think from a command line POV, it is  
more consistent to be able to do 'hg revert subrepo/dir' or 'hg revert  
subrepo/file'.  A subrepo clean update prevents that (or at the very least  
is inconsistent with that, if we keep the current behavior as well when no  
path or only the root of the subrepo is given.)

Any other top level repo using the subrepo is still in a consistent state,  
until it updates the subrepo to a different revision.  (And then it is  
free to add more commits to the subrepo to "fix" it before committing a  
specific subrepo rev).

The reason I brought up backout is because in a flat repo, a one line  
commit even hundreds of revs back can be surgically nuked.  But a 'backout  
-S' using the current revert would kill *all* of the subsequent changes,  
and I think that would be very surprising.

I remember the threads a few years back about 'should commit recurse into  
subrepos or abort?'.  You argued consistency then, and my take away was  
that letting a user accidentally commit something not identical to the  
filesystem would be a nasty (and permanent, without history editing)  
surprise.  I don't see how that principle applies here either though-  
revert isn't recording history.

>> It's been like this for 3 years now, but revert subrepo support is  
>> totally
>> undocumented AFAICT- nothing in 'hg help revert' or in 'hg help
>> subrepos'.  Would you consider this a bug (and accept patches to fix
>> this), or are we stuck with this?
> If "been like this" means we do get clean checkouts today, then I think
> we should stick with it.

Yes.  The current code basically translates the parent rev to a subrepo  
rev, and does 'hg update -C subrev', after making the backup files.


More information about the Mercurial-devel mailing list