On 13/06/2019 17:24, Jeff King wrote:
> On Thu, Jun 13, 2019 at 09:05:16AM -0700, Junio C Hamano wrote:
>
>> aleksandrs@xxxxxxxxxxxx writes:
>>
>>> My repo indeed contains a ".git/sequencer/todo" file which
>>> contains references to commits long-gone (i.e., rebased).
>>> Renaming or deleting this file stops whines about "error: could
>>> not parse".
I think this bug report is a result of
4a72486de9 ("fix cherry-pick/revert status after commit", 2019-04-16)
which tries to read the sequencer todo file if it is present and there
is no CHERRY_PICK_HEAD/REVERT_HEAD. Before that if the user committed a
conflict resolution in the middle of a sequence of picks then status
would forget that there was a cherry-pick in progress. It just reuses
the code that cherry-pick uses to parse the first todo list item which
expects the commits to exist.
>> Interesting. So in short, when the repository has leftover
>> sequencer state file that is not in use, "git status parse" thing
>> (whatever it is---are you getting it when you run "git status"
>> command???)---is not careful enough to notice that it does not
>> matter even if that leftover file is unusable.
>>
>> Two issues "the sequencer" folks may want to address are
>>
>> (1) make the one that reads an irrelevant/stale 'todo' file more
>> careful to ignore errors in such a file;
>>
>> (2) make the sequencer machinery more careful to clean up after it
>> is done or it is aborted (for example, "git reset --hard"
>> could remove these state files preemptively even when a rebase
>> is not in progress, I would think).
>>
>> I think we already had some patches toward the latter recently.
>
> Maybe I am not understanding it correctly, but do you mean in (2) that
> "git reset --hard" would always clear sequencer state?
I think the commit Junio is referring to is
b07d9bfd17 ("commit/reset: try to clean up sequencer state", 2019-04-16)
which will only remove the sequencer directory if it stops after the
pick was the last one in the series. The idea is that if cherry-pick
stops for a conflict resolution on the last pick user commits the result
directly or run reset without running `cherry-pick --continue`
afterwards the sequencer state gets cleaned up properly.
> That seems
> undesirable to me, as I often use "git reset" to move the head around
> during a rebase. (e.g., when using "rebase -i" to split apart I commit,
> I stop on that commit, "git reset" back to the parent, and then
> selectively "add -p" the various parts).
>
> Direction (1) seems quite sensible to me, though.
Now that we try harder to clean up the sequencer state I wonder if that
would just cover up bugs where the state has not been removed when it
should have been. That can lead to unpleasant problems if the user
aborts a single revert/cherry-pick when there is stale sequencer state
around as it rewinds HEAD to the commit when the stale
cherry-pick/revert was started as explained in the message to b07d9bfd17
("commit/reset: try to clean up sequencer state", 2019-04-16)
If we do want to do something then maybe teaching gc not to collect
commits listed in .git/sequencer/todo and
.git/rebase-merge/git-rebase-todo would be useful.
Best Wishes
Phillip
> -Peff
>
