Hi Oswald
On 26/04/2023 16:34, Oswald Buddenhagen wrote:
On Sun, Mar 26, 2023 at 03:28:01PM +0100, Phillip Wood wrote:
On 23/03/2023 16:22, Oswald Buddenhagen wrote:
When the user butchers the todo file during rebase -i setup, the
--continue which would follow --edit-todo would have skipped the last
steps of the setup. Notably, this would bypass the fast-forward over
untouched picks (though the actual picking loop would still fast-forward
the commits, one by one).
Fix this by splitting off the tail of complete_action() to a new
start_rebase() function and call that from sequencer_continue() when no
commands have been executed yet.
More or less as a side effect, we no longer checkout `onto` before
exiting
when the todo file is bad.
I think the implications of this change deserve to be discussed in the
commit message. Three things spring to mind but there may be others I
haven't thought of
- Previously when rebase stopped and handed control back to the user
HEAD would have already been detached. This patch changes that
meaning we can have an active rebase of a branch while that branch is
checked out. What does "git status" show in this case? What does the
shell prompt show? Will it confuse users?
the failed state is identical to the "still editing the initial todo"
state as far as "git status" and the shell prompt are concerned. this
seems reasonable. i'll add it to the commit message.
When you do that please mention what "git status" and the shell prompt
actually print in this case. Ideally "git status" should mention that
the todo list needs to be edited if there are still errors in it, though
it would not surprise me if it is not that helpful at the moment.
- Previously if the user created a commit before running "rebase
--continue" we'd rebase on to that commit. Now that commit will be
silently dropped.
this is arguably a problem, but not much different from the pre-existing
behavior of changes to HEAD done during the initial todo edit being lost.
I think there is a significant difference in that we're moving from a
situation where we lose commits that are created while rebase is running
to one where we're losing commits created while rebase is stopped. If a
user tries to create a commit while rebase is running then they're
asking for trouble. I don't think creating commits when rebase is
stopped is unreasonable in the same way.
to avoid that, we'd need to lock HEAD while editing the todo. is that
realistic at all?
I don't think it is practical to lock HEAD while git is not running. We
could just check HEAD has not changed when the rebase continues after
the user has fixed the todo list as you suggest below.
on top of that, i should verify HEAD against orig-head in
start_rebase(). though the only way for the user to get out of that
situation is saving the todo contents and --abort'ing (and we must take
care not the touch HEAD).
I think in that case it wouldn't be terrible to lose the edited todo
list as it is a bit of a corner case. The simplest thing to do would be
to print an error and remove .git/rebase-merge.
this is somewhat similar to the abysmal situation of the final
update-ref failing if the target ref has been modified while being
rebased. we'd need to lock that ref for the entire duration of the
rebase to avoid that.
"abysmal" is rather harsh - it would also be bad to overwrite the ref in
that case. I think it in relatively hard to get into that situation
though as "git checkout" wont checkout a branch that is being updated by
a rebase.
- Previously if the user checkout out another commit before running
"rebase --continue" we'd rebase on to that commit. Now we we rebase
on to the original "onto" commit.
this can be subsumed into the above case.
Meaning check and error out if HEAD has changed?
> This makes aborting cheaper and will simplify
> things in a later change.
Given that we're stopping so the user can fix the problem and continue
the rebase I don't think optimizing for aborting is a convincing
reason for this change on its own.
this is all part of the "More or less as a side effect" paragraph, so
this isn't a relevant objection.
I'm simply saying that we should not be optimizing for "rebase --abort"
in this case. Do you think we should?
diff --git a/builtin/revert.c b/builtin/revert.c
index 62986a7b1b..00d3e19c62 100644
--- a/builtin/revert.c
+++ b/builtin/revert.c
@@ -231,7 +231,8 @@ static int run_sequencer(int argc, const char
**argv, struct replay_opts *opts)
return ret;
}
if (cmd == 'c')
- return sequencer_continue(the_repository, opts);
+ return sequencer_continue(the_repository, opts,
+ 0, NULL, NULL, NULL);
It's a bit unfortunate that we have to start passing all these extra
parameters, could the sequencer read them itself in read_populate_opts()?
that wouldn't help in this case, as these are dummy values which aren't
going to be used.
If we only need to pass these when rebasing maybe we should have
separate wrappers for continuing a rebase and a cherry-pick/revert. If
we don't always need these parameters when continuing a rebase we could
have a separate function when these parameters are required and leave
the signature of sequencer_continue() unchanged.
but more broadly, the whole state management is a total mess.
For historic reasons there are separate functions to write the state for
the "merge" backed and the "apply" backend. That is not ideal but it is
hardly a "total mess". The code for reading the state files is more
contrived than the code that writes them. I do have some patches to try
and reduce the duplication when reading the state files.
-int sequencer_continue(struct repository *r, struct replay_opts *opts)
+static int start_rebase(struct repository *r, struct replay_opts
*opts, unsigned flags,
+ const char *onto_name, const struct object_id *onto,
+ const struct object_id *orig_head, struct todo_list
*todo_list);
It would be nice to avoid this forward declaration. I think you could
do that by adding a preparatory patch that moves either
checkout_onto() or sequencer_continue()
i went for the "minimal churn" approach.
There is a balance to be had, but we don't want to build up a lot of
forward declarations over time just because it is easier than moving the
function in a preparatory patch. A simple patch to move a function is
easy to review with --color-moved.
+ git reflog > reflog &&
+ test $(grep -c fast-forward reflog) = 1 &&
Using test_line_count would make test failures easier to debug.
that's calling for a new test_filtered_line_count function which would
have quite some users.
for the time being, both grep + test_line_count and grep -c are rather
prevalent, in this file the latter in particular.
The style of our tests has evolved over time. When adding new tests it
is better to focus on making them easy to debug. I don't think you need
to add a new function here, just
grep fast-forward reflog >filtered-reflog
test_line_count = 1 filtered-reflog
+ test_cmp_rev HEAD~1 primary~1 &&
+ test "$(git log -1 --format=%B)" = "E_reworded"
It is slightly more work, but please use test_cmp for things like this
as it makes it so much easier to debug test failures.
fair enough, but the precedents again speak a different language.
Yes older tests tend to be written in a style that is harder to debug.
Best Wishes
Phillip
