Re: [PATCH 1/2] Move code resolving packed refs into its own function.

[Junio C Hamano <junkio@cox.net>]

Christian Couder <chriscool@tuxfamily.org> writes:

> Junio C Hamano wrote:
>> Christian Couder <chriscool@tuxfamily.org> writes:
>> > This patch move Linus' packed refs resolving code from
>> > "resolve_ref" into a new "resolve_packed_ref" extern
>> > function so that it can be reused when needed.
>>
>> I think we are stepping on each other's toes.  How far into the
>> process of making correct branch deletion are you?
>
> I am not farther than the 2 patches I sent yesterday (before going to bed).

Thanks.  I just did not want to waste your work with overlapping
duplicated efforts.

I think what we have in "next" tonight is in a more-or-less
testable shape, although it has still a long way to reach
"master".  Things I know we need to address:

 - I've updated lock_ref_sha1_basic() to remove empty left-over
   directories and to notice conflicts between 'foo/bar' vs
   'foo' when creating a new ref, hopefully in the same spirit
   as your patch to safe_create_leading_directories(), but done
   differently (safe_... function is meant to be used anywhere
   not just $GIT_DIR/refs/, and it felt wrong for it to take
   exception for packed refs).  We should do the same for the
   reflog hierarchy but we currently don't.

 - We need to audit our shell scripts to make sure they do not
   depend on being able to look directly into $GIT_DIR/refs to
   see if the ref they are interested in exists.  I've fixed a
   few in git-fetch while handling the patch to clean up its
   output from Santi, but I would not be surprised if there are
   more.  The code in git-branch and git-tag to list what's
   there are Ok; they use "rev-parse --symbolic --all/--tags".

 - I think gitweb should be Ok; it does peek-remote on the
   repository.  Although we would probably want to update
   git_get_references and git_get_refs_list sub to use
   for-each-ref there, that can be done as a later optimization.

 - Dumb transports are not aware of packed refs on the remote
   side.  The underlying commit walkers (anything that links
   with fetch.c) needs their fetch_ref() implementation updated
   to look at the packed-refs file from the remote side and we
   should be fine after that.  I haven't looked at rsync
   transport but the change necessary there shouldn't be too
   involved.