diff mbox series

Re* [PATCH v2] clone: Allow combining --bare and --origin

Message ID xmqq4kc0j4cd.fsf_-_@gitster.g (mailing list archive)
State New, archived
Headers show
Series Re* [PATCH v2] clone: Allow combining --bare and --origin | expand

Commit Message

Junio C Hamano Aug. 7, 2021, 10:08 p.m. UTC 
Junio C Hamano <gitster@pobox.com> writes:

>>> It is somewhat unfortunate that we do not say what the name of the
>>> "origin" is anywhere in the resulting configuration file.  The only
>>> way to tell that "--origin somewhere" was used is to notice that there
>>> is only one remote and its name is "somewhere".
>> ...
> But we'd end up treating them the same.  And something like
> remote.originName would help that.  Otherwise, we'd end up sending
> this message:
>
>     Even if we give "--bare --origin yourfavouritename" to you now,
>     unlike how 'origin' is treated in the default case, in the
>     resulting repository, 'yourfavouritename' is not special at all.
>
> Some people may want to treat yourfavouritename is not special at
> all, while some people may want to treat yourfavouritename truly as
> a replacement for 'origin' that is the default.  The message we
> would be sending is that we'd ignore the latter folks.

So, let's illustrate one of the things that is needed after the good
first step to allow --bare --origin=yourfavouritename used together.

There may be other things that needs fixing, of course, but we need
to start from somewhere.

---- >8 -------- >8 -------- >8 -------- >8 -------- >8 -------- >8 ----
Subject: [PATCH] remote: fall back on the sole remote when unspecified

Historically, we used hardcoded "origin" as the fallback default for
commands that take a remote (e.g. "git fetch") when the user did not
tell us otherwise.  Since the "--origin=name" option was taught to
"git clone", however, we may not have a remote whose name is
"origin" at all.

Which means that the name given to "git clone --origin" does not
truly replace the hardcoded "origin". An example of such limitation
is that "git fetch" (no other parameters) would fetch happily from
the "origin" repository, but in a repository cloned with the custom
name using "--origin=name", "git fetch" would not fetch from anywhere
and instead fail.

We can fix this by noticing that the repository has one and only one
remote defined, and use that as a replacement for the hardcoded
"origin".

Signed-off-by: Junio C Hamano <gitster@pobox.com>
---

This matters for automation for those who want to use --origin
option.  Imagine you have multiple bare clones and you wanted to use
custom names for 'origin'.  And you want a cron job that goes over
these repositories and run "git fetch" from their upstream before
you come in for work, so that these bare clones can be used as
close-by mirrors of their upstream projects.

Unfortunately, that would not work.  If these repositories use
their own nicknames for their upstream that are not "origin",

	for repo in a b c
	do
		git -C $repo fetch
	done

would just fail.  Of course, you can somehow out-of-band know the
origin's name for each repo, e.g.

	for repoorigin in a:xyzzy b:frotz c:nitfol
	do
		repo=${repoorigin%:*}
                origin=${repoorigin#*:}
		git -C $repo fetch $origin
	done

but that is solving a problem that arises only because we are not
treating the name given to "git clone --origin=name" as a true
replacement for the default "origin".

 remote.c             | 10 +++++++++-
 t/t5512-ls-remote.sh | 10 ++++++++--
 2 files changed, 17 insertions(+), 3 deletions(-)

Comments

Roman Neuhauser Aug. 8, 2021, 2:03 a.m. UTC | #1 
# gitster@pobox.com / 2021-08-07 15:08:02 -0700:
> Subject: [PATCH] remote: fall back on the sole remote when unspecified
> 
> Historically, we used hardcoded "origin" as the fallback default for
> commands that take a remote (e.g. "git fetch") when the user did not
> tell us otherwise.  Since the "--origin=name" option was taught to
> "git clone", however, we may not have a remote whose name is
> "origin" at all.
> 
> Which means that the name given to "git clone --origin" does not
> truly replace the hardcoded "origin". An example of such limitation
> is that "git fetch" (no other parameters) would fetch happily from
> the "origin" repository, but in a repository cloned with the custom
> name using "--origin=name", "git fetch" would not fetch from anywhere
> and instead fail.

hey, i'm all for all this pre-existing lossage getting fixed if you
can do it.  all i'm saying is that since this combination of options
wasn't possible before there won't be any pre-existing uses of git
suddenly breaking.

> This matters for automation for those who want to use --origin
> option.  Imagine you have multiple bare clones and you wanted to use
> custom names for 'origin'.  And you want a cron job that goes over
> these repositories and run "git fetch" from their upstream before
> you come in for work, so that these bare clones can be used as
> close-by mirrors of their upstream projects.

imagine that you wanted to use git clone --bare --origin with
any git version released so far.  this is not snark, i'm pointing
out that git git has a history of things not working where one
would expect them to.
 
> Unfortunately, that would not work.  If these repositories use
> their own nicknames for their upstream that are not "origin",
> 
> 	for repo in a b c
> 	do
> 		git -C $repo fetch
> 	done

  for repo in a b c; do
    git -C $repo fetch --all # or git -C remote update
  done

all it takes to mitigate this is to point this out in the release
notes and man page.  what you sketched out above is analogous to my
initial encounter with git clone --bare --origin not working:
where were you when the half-assed implementation was landing?  :)
why was there no one to champion for people who'd want to use those
two together? :)) (j/k)
 
> would just fail.  Of course, you can somehow out-of-band know the
> origin's name for each repo, e.g.

even if i accept the premise that git fetch --all can't be used
and the explicit name is necessary, isn't that magical out-of-band
wand called git-config?

  origin=$(
    git config --file $repo \
    --name-only --get-regexp \
    '^remote\.[^.]*.url' |
    sed -E 's/^remote\.([^.]+).url$/\1/'
  )
  git -C $repo fetch $origin
 
i'm not skilled enough in git-config to simplify that.


i think it'd be prudent to pause this thread for now because it's
only distracting you from fixing the --origin fallout, and as long
as you talk about how it *should* be while i bring up available
workarounds, it's just noise.

> but that is solving a problem that arises only because we are not
> treating the name given to "git clone --origin=name" as a true
> replacement for the default "origin".

and i'm really grateful that you're tying the loose ends, as long
as this whole thing doesn't fizzle out on account of being too much,
and the partial improvement doesn't get swept with it!
 
i think i said in earlier that i'm a big fan of stripping "origin"
of its special standing.  huge kudos if you can see this through.
diff mbox series

Patch

diff --git c/remote.c w/remote.c
index dfb863d808..8a2fd1ccc9 100644
--- c/remote.c
+++ w/remote.c
@@ -39,6 +39,8 @@  static int remotes_alloc;
 static int remotes_nr;
 static struct hashmap remotes_hash;
 
+static const char *default_remote_name;
+
 static struct branch **branches;
 static int branches_alloc;
 static int branches_nr;
@@ -460,6 +462,12 @@  static void read_config(void)
 	}
 	git_config(handle_config, NULL);
 	alias_all_urls();
+	if (remotes_nr == 1 &&
+	    remotes[0]->configured_in_repo &&
+	    remotes[0]->url)
+		default_remote_name = remotes[0]->name;
+	else
+		default_remote_name = "origin";
 }
 
 static int valid_remote_nick(const char *name)
@@ -483,7 +491,7 @@  const char *remote_for_branch(struct branch *branch, int *explicit)
 	}
 	if (explicit)
 		*explicit = 0;
-	return "origin";
+	return default_remote_name;
 }
 
 const char *pushremote_for_branch(struct branch *branch, int *explicit)
diff --git c/t/t5512-ls-remote.sh w/t/t5512-ls-remote.sh
index f53f58895a..aa6f14e8fd 100755
--- c/t/t5512-ls-remote.sh
+++ w/t/t5512-ls-remote.sh
@@ -83,8 +83,14 @@  test_expect_success 'ls-remote --sort="-refname" --tags self' '
 	test_cmp expect actual
 '
 
-test_expect_success 'dies when no remote specified and no default remotes found' '
-	test_must_fail git ls-remote
+test_expect_success 'ls-remote falls back to the only remote' '
+	generate_references \
+		refs/tags/mark1.2 \
+		refs/tags/mark1.10 \
+		refs/tags/mark1.1 \
+		refs/tags/mark >expect &&
+	git ls-remote --sort="-refname" --tags >actual &&
+	test_cmp expect actual
 '
 
 test_expect_success 'use "origin" when no remote specified' '