Re: [BUG] Git fast-export with import marks file omits merge commits

["Martin Ågren" <martin.agren@gmail.com>]

On 20 April 2018 at 00:48, Junio C Hamano <gitster@pobox.com> wrote:
> Isaac Chou <Isaac.Chou@microfocus.com> writes:
>
>> I inspected the source code (builtin/fast-export.c) for the
>> fast-export issue I encountered, and it looks like the merge
>> commit is discarded too early by the call to object_array_pop()
>> after only one of the two UNSHOWN parents is processed in the
>> method handle_tail().  The poped merge commit still has one
>> UNSHOWN parent, therefore it is not processed and is lost in the
>> output.  Can someone advise me on how to submit a code change or
>> bug report in order to get the fix into the code base?
>
> There indeed are some differences between v2.14 and v2.15 around the
> code that returns early when has_unshown_parent() says "yes" [*1*],
> but the decision to return early when the function says "yes" hasn't
> changed between that timeperiod---it dates back to f2dc849e ("Add
> 'git fast-export', the sister of 'git fast-import'", 2007-12-02),
> i.e. the very beginning of the program's life.
>
> I'll CC those who wrote the original and b3e8ca89 ("fast-export: do
> not copy from modified file", 2017-09-20) and 71992039
> ("object_array: add and use `object_array_pop()`", 2017-09-23),
> which are the only two commits that touch the surrounding area
> during that timeperiod, to ask if something jumps at them.
>
> Thanks.
>
>
> [Footnotes]
>
> *1* An excerpt from 'git diff v2.14.0 v2.15.0 builtin/fast-export.c'
>     reads like so:
>
> diff --git a/builtin/fast-export.c b/builtin/fast-export.c
> index d412c0a8f3..2fb60d6d48 100644
> --- a/builtin/fast-export.c
> +++ b/builtin/fast-export.c
> ...
> @@ -630,15 +645,15 @@ static void *anonymize_tag(const void *old, size_t *len)
>         return strbuf_detach(&out, len);
>  }
>
> -static void handle_tail(struct object_array *commits, struct rev_info *revs)
> +static void handle_tail(struct object_array *commits, struct rev_info *revs,
> +                       struct string_list *paths_of_changed_objects)
>  {
>         struct commit *commit;
>         while (commits->nr) {
> -               commit = (struct commit *)commits->objects[commits->nr - 1].item;
> +               commit = (struct commit *)object_array_pop(commits);
>                 if (has_unshown_parent(commit))
>                         return;
> -               handle_commit(commit, revs);
> -               commits->nr--;
> +               handle_commit(commit, revs, paths_of_changed_objects);
>         }
>  }

Indeed. This looks wrong and the guilty person would be me.

If my 71992039 ("object_array: add and use `object_array_pop()`",
2017-09-23) would instead have done something like
s/commits->nr--/(void)object_array_pop(commits)/ it would not have
screwed up as much. It could also use a peek+pop-pattern.

Isaac, are you up for submitting a patch? Just let me know if you
encounter any issues. Otherwise, I can submit a patch shortly since I
was the one who dropped the ball originally.

Thanks for diagnosing this.

Martin