Re: [PATCH] rebase: set REF_HEAD_DETACH in checkout_up_to_date()

From: John Cai <johncai86@gmail.com>
Date: Mar 11, 2022, 14:14 UTC
Message-ID: <B1A94338-404E-43F2-AF3C-889097FD938B@gmail.com>
In-Reply-To: <xmqq5yolav8l.fsf@gitster.g>

Hi Junio,

On 11 Mar 2022, at 0:55, Junio C Hamano wrote:

Show 30 quoted lines

> "John Cai via GitGitGadget" <gitgitgadget@gmail.com> writes:
>
>> This is happening because on a fast foward with an oid as a <branch>,
>> update_refs() will only call update_ref() with REF_NO_DEREF if
>> RESET_HEAD_DETACH is set. This change was made in 176f5d96 (built-in rebase
>> --autostash: leave the current branch alone if possible,
>> 2018-11-07). In rebase, we are not setting the RESET_HEAD_DETACH flag,
>> which means that the update_ref() call ends up dereferencing
>> HEAD and updating it to the oid used as <branch>.
>>
>> The correct behavior is that git rebase should update HEAD to $(git
>> rev-parse topic) without dereferencing it.
>
> It is unintuitive that unconditionally setting the RESET_HEAD_DETACH
> bit is the right solution.
>
> If the command weren't "rebase master side^0" but "rebase master
> side", i.e. "please rebase the side branch itself, not an unnamed
> branch created out of the side branch, on master", according to
> <reset.h>, we ought to end up being on a detached HEAD, as
> reset_head() with the bit
>
>     /* Request a detached checkout */
>     #define RESET_HEAD_DETACH (1<<0)
>
> requests a detached checkout.  But that apparently is not what would
> happen with your patch applied.
>
> Puzzled.  The solution to the puzzle probably deserves to be in the
> proposed log message.

Good point. Thinking aloud, here is the callstack.

checkout_up_to_date() -> reset_head() -> update_refs() -> update_ref()

if <branch> is not a valid ref, rebase_options head_name is set to NULL. This eventually leads update_refs() to decide that it doesn't need to switch to a branch via its switch_to_branch variable.

reset.c:

if (!switch_to_branch)
	ret = update_ref(reflog_head, "HEAD", oid, head,
			 detach_head ? REF_NO_DEREF : 0,
			 UPDATE_REFS_MSG_ON_ERR);
 else {
	ret = update_ref(reflog_branch ? reflog_branch : reflog_head,
			 switch_to_branch, oid, NULL, 0,
			 UPDATE_REFS_MSG_ON_ERR);
	if (!ret)
		ret = create_symref("HEAD", switch_to_branch,
				    reflog_head);
}

since the flags do not include RESET_HEAD_DETACH, detach_head is set to false and we get a deferenced HEAD update.

The solution I came up with works because when <branch> __is__ a valid branch, udpate_refs() takes a different code path that calls create_symref() with a branch, which is why we don't end up with a detached HEAD.

I see why this is confusing though. From checkout_up_to_date's perspective it looks like we are unconditionally detaching HEAD. So what we could do is only set the flag in checkout_up_to_date() when, from checkout_up_to_date's perspective, we will end up with a detached head. something like this:

diff --git a/builtin/rebase.c b/builtin/rebase.c
index b29ad2b65e72..f0403fb12421 100644
--- a/builtin/rebase.c
+++ b/builtin/rebase.c
@@ -827,8 +827,10 @@ static int checkout_up_to_date(struct rebase_options *options)
                    getenv(GIT_REFLOG_ACTION_ENVIRONMENT),
                    options->switch_to);
        ropts.oid = &options->orig_head;
-       ropts.branch = options->head_name;
        ropts.flags = RESET_HEAD_RUN_POST_CHECKOUT_HOOK;
+       ropts.branch = options->head_name;
+       if (!ropts.branch)
+               ropts.flags |=  RESET_HEAD_DETACH;
        ropts.head_msg = buf.buf;
        if (reset_head(the_repository, &ropts) < 0)
                ret = error(_("could not switch to %s"), options->switch_to);

Otherwise, checkout_up_to_date() has to implicitly know the downstream logic in
update_refs(). I believe that's the main source of the confusion--is that right?

>
> Thanks.

Thanks
John