Re: [PATCH v2 0/6] Fix checkout-dash to work with rebase

Ramkumar Ramachandra [off-list ref] writes:

So after extensive discussions with Junio, I have updated [5/6] to
special-case rebase and rebase -i instead of dropping the "HEAD
detached from" message altogether.  Also, [1/6] includes two more
tests, as suggested by Junio.

Junio: The message is now the constant "rebase in progress; onto
$ONTO".

That message is better than I was anticipating.

I was actually assuming that you would leave them as they are
i.e. "# HEAD detached at xxx", as we agreed that we will see them
improved anyway by the other topic.

I am still puzzled to see that the change to checkout comes at the
very end.  With "do not depend on rebase reflog messages" done
before the "checkout: respect reflog-action", I was hoping that we
can have changes to the actual reflog message made to rebase (patch
2 and 3) can be done at the very end.  Was I missing something else?

In other words, the order I was anticipating to see after the
discussion (this is different from saying "A series that is not
ordered like this is unacceptable") was:

  wt-status: remove unused field in grab_1st_switch_cbdata

This is an unrelated clean-up, and can be done before anything else.

  t/checkout-last: checkout - doesn't work after rebase

This spells out what we want to happen at the end and marks the
current breakage.

  status: do not depend on rebase reflog messages

This compensates for fallouts from the next change.

  checkout: respect GIT_REFLOG_ACTION

And this is the fix, the most important step.

  rebase: prepare to write reflog message for checkout
  rebase -i: prepare to write reflog message for checkout

And these are icing on the cake, but that cannot be done before
status is fixed.

Junio C Hamano [off-list ref] writes:

In other words, the order I was anticipating to see after the
discussion (this is different from saying "A series that is not
ordered like this is unacceptable") was:

quoted

  wt-status: remove unused field in grab_1st_switch_cbdata

This is an unrelated clean-up, and can be done before anything else.

quoted

  t/checkout-last: checkout - doesn't work after rebase

This spells out what we want to happen at the end and marks the
current breakage.

quoted

  status: do not depend on rebase reflog messages

This compensates for fallouts from the next change.

quoted

  checkout: respect GIT_REFLOG_ACTION

And this is the fix, the most important step.

quoted

  rebase: prepare to write reflog message for checkout
  rebase -i: prepare to write reflog message for checkout

And these are icing on the cake, but that cannot be done before
status is fixed.

I actually tried to reorder the patches and they seem to work OK as
expected.  And I think it makes sense to order them the way I've
been suggesting, so I'll tentatively queue the result of reordering
on 'rr/rebase-checkout-reflog' and push it out as a part of 'pu'.

Please check to see if I introduced a new bug while doing so.

Regardless of the ordering, however, I suspect two patches that
change the message recorded in reflog in "rebase" need further
fixing.

For example, the one in "git reabse" does this:

    GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name"
    git checkout -q "$onto^0" || die "could not detach HEAD"
    git update-ref ORIG_HEAD $orig_head
    ...
    run_specific_rebase

But the specific rebase, e.g. git-rebase--interactive, does this:

	case $head_name in
	refs/*)
		message="$GIT_REFLOG_ACTION: $head_name onto $onto" &&
		git update-ref -m "$message" $head_name $newhead $orig_head &&
		git symbolic-ref \
		  -m "$GIT_REFLOG_ACTION: returning to $head_name" \
		  HEAD $head_name
		;;
	esac && {

I think the message you added to "git reabse" is only meant for that
specific "checkout $onto", but it is set globally.  Wouldn't it
affect later use, which expected it to be "rebase" and nothing else?

Perhaps something like this on top of the entire series may be
sufficient (which will be queued as "SQUASH???" at the tip).

Hint for anybody (not limited to Ram):

There also are other 'git checkout' invocations in git-rebase\*.sh
that are not yet covered by these "nicer reflog message" patches,
which may want to be fixed up before these two icing-on-cake patches
become ready to be merged; see output of

	git grep -C2 'git checkout' -- git-rebase\*.sh

Thanks.

 git-rebase--interactive.sh    | 15 ++++++++++-----
 git-rebase.sh                 |  4 ++--
 t/t3404-rebase-interactive.sh | 15 +++++++++++++++
 3 files changed, 27 insertions(+), 7 deletions(-)

diff --git a/git-rebase--interactive.sh b/git-rebase--interactive.sh
index a05a6e4..f21ff7c 100644
--- a/git-rebase--interactive.sh
+++ b/git-rebase--interactive.sh

@@ -837,9 +837,11 @@ comment_for_reflog start
 
 if test ! -z "$switch_to"
 then
-	GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $switch_to"
-	output git checkout "$switch_to" -- ||
-		die "Could not checkout $switch_to"
+	(
+		GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $switch_to"
+		export GIT_REFLOG_ACTION
+		output git checkout "$switch_to" --
+	) || die "Could not checkout $switch_to"
 fi
 
 orig_head=$(git rev-parse --verify HEAD) || die "No HEAD?"

@@ -981,7 +983,10 @@ has_action "$todo" ||
 
 test -d "$rewritten" || test -n "$force_rebase" || skip_unnecessary_picks
 
-GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name"
-output git checkout $onto || die_abort "could not detach HEAD"
+(
+	GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name"
+	export GIT_REFLOG_ACTION
+	output git checkout $onto
+) || die_abort "could not detach HEAD"
 git update-ref ORIG_HEAD $orig_head
 do_rest

diff --git a/git-rebase.sh b/git-rebase.sh
index 7d55372..2344eef 100755
--- a/git-rebase.sh
+++ b/git-rebase.sh

@@ -523,8 +523,8 @@ test "$type" = interactive && run_specific_rebase
 # Detach HEAD and reset the tree
 say "$(gettext "First, rewinding head to replay your work on top of it...")"
 
-GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name"
-git checkout -q "$onto^0" || die "could not detach HEAD"
+GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name" \
+	git checkout -q "$onto^0" || die "could not detach HEAD"
 git update-ref ORIG_HEAD $orig_head
 
 # If the $onto is a proper descendant of the tip of the branch, then

diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
index a58406d..c175ef1 100755
--- a/t/t3404-rebase-interactive.sh
+++ b/t/t3404-rebase-interactive.sh

@@ -934,6 +934,21 @@ test_expect_success 'rebase --edit-todo can be used to modify todo' '
 	test L = $(git cat-file commit HEAD | sed -ne \$p)
 '
 
+test_expect_success 'rebase -i produces readable reflog' '
+	git reset --hard &&
+	git branch -f branch1 H &&
+	git rebase -i --onto I F branch1 &&
+	cat >expect <<-\EOF &&
+	rebase -i (start): checkout I
+	rebase -i (pick): G
+	rebase -i (pick): H
+	rebase -i (finish): returning to refs/heads/branch1
+	EOF
+	tail -n 4 .git/logs/HEAD |
+	sed -e "s/.*	//" >actual &&
+	test_cmp expect actual
+'
+
 test_expect_success 'rebase -i respects core.commentchar' '
 	git reset --hard &&
 	git checkout E^0 &&

Junio C Hamano wrote:

I actually tried to reorder the patches and they seem to work OK as
expected.  And I think it makes sense to order them the way I've
been suggesting, so I'll tentatively queue the result of reordering
on 'rr/rebase-checkout-reflog' and push it out as a part of 'pu'.

Please check to see if I introduced a new bug while doing so.

Thanks for the reorder and commit message tweaks.  I'm working on the
series you put up on `pu` now.

For example, the one in "git reabse" does this:

    GIT_REFLOG_ACTION="$GIT_REFLOG_ACTION: checkout $onto_name"
    git checkout -q "$onto^0" || die "could not detach HEAD"
    git update-ref ORIG_HEAD $orig_head
    ...
    run_specific_rebase

But the specific rebase, e.g. git-rebase--interactive, does this:

        case $head_name in
        refs/*)
                message="$GIT_REFLOG_ACTION: $head_name onto $onto" &&
                git update-ref -m "$message" $head_name $newhead $orig_head &&
                git symbolic-ref \
                  -m "$GIT_REFLOG_ACTION: returning to $head_name" \
                  HEAD $head_name
                ;;
        esac && {

I think the message you added to "git reabse" is only meant for that
specific "checkout $onto", but it is set globally.  Wouldn't it
affect later use, which expected it to be "rebase" and nothing else?

Both rebase.sh and rebase--interactive.sh set a sane GIT_REFLOG_ACTION
right on top (using set_reflog_action), so no worries.  I'll just
double-check to make sure that no bogus/ incorrect messages are
printed.

Perhaps something like this on top of the entire series may be
sufficient (which will be queued as "SQUASH???" at the tip).

I think this takes the wrong approach to the problem.  In my opinion,
the correct approach is to actually overshadow die() with a function
that clears GIT_REFLOG_ACTION before calling die().

        git grep -C2 'git checkout' -- git-rebase\*.sh

Ugh.  I'll check all the codepaths thoroughly before submitting a re-roll.

Thanks.

Ramkumar Ramachandra wrote:

Thanks for the reorder and commit message tweaks.  I'm working on the
series you put up on `pu` now.

That said, I do not agree with one of your commit message updates:

    checkout: respect GIT_REFLOG_ACTION

    GIT_REFLOG_ACTION is an environment variable specifying the reflog
    message to write after an action is completed.  Several other commands
    including merge, reset, and commit respect it.

    Fix the failing tests in t/checkout-last by making checkout respect it
    too.  You can now expect

      $ git checkout -

    to work as expected after any operation that internally uses "checkout"
    as its implementation detail, e.g. "rebase".

After the patch you _cannot_ expect checkout-dash to work after any
operation that internally uses checkout; it's limited to rebase.  Many
other scripts (eg. bisect) do _not_ set GIT_REFLOG_ACTION at all.

Junio C Hamano wrote:

I don't know if this test is a good idea at all.  We never directly
check the messages written by a command to the reflog, and I suspect
that there is a good reason for this: the format of .git/logs is not
guaranteed to be stable, and the messages written by various commands
are not guaranteed to be stable either; the only machine-parsing of
reflogs we do is very minimal: interpret_nth_prior_checkout() and
grab_1st_switch().

A quick

  $ git grep .git/logs -- t

shows that I'm mostly right about this.

Why make an exception in the case of rebase -i?  In the worst case,
the patch atleast needs to be discussed as an independent patch: it's
certainly not obvious enough to sneak into this series.  I'll submit a
re-roll without this hunk.