Re: [BUG] v2.16.0-rc0 seg faults when git bisect skip

[Yasushi SHOJI <yasushi.shoji@gmail.com>]

Hi Martin,

Thank you for your comment.
I haven't have time to read the code carefully so bare with me.

On Sat, Jan 6, 2018 at 5:21 PM, Martin Ågren <martin.agren@gmail.com> wrote:
>> On Fri, Jan 5, 2018 at 11:45 AM, Yasushi SHOJI <yasushi.shoji@gmail.com> wrote:
>>> When does the list allowed to contain NULLs?
>
> Short answer: there are no commits left to test.
>
> The list is built in the for-loop in `find_bisection()`. So the
> technical answer is: if all commits in the initial list `commit_list`
> are UNINTERESTING (including if `commit_list` is empty to begin with).
>
> It's also helpful to study where we should end up from there. We should
> take the `if (!revs.commits)` branch in `bisect_next_all()`. That is, we
> should print either "There are only 'skip'ped commits left to test. The
> first bad commit could be any of:" or "<commit> was both good and bad".

best_bisection_sorted() seems to do

 - get the commit list along with the number of elements in the list
 - walk the list one by one to check whether a element have TREESAME or not
 - if TREESAME, skip
 - if not, add it to array
 - sort the array by distance
 - put elements back to the list

so, if you find TREESAME, you get less elements than given, right?
Also, if you sort, the last commit, which has NULL in the ->next,
might get in the middle of the array??

# BTW, is it really fast to use QSORT even if you have to convert to
# an array from list?

>>> Since nobody noticed it since 7c117184d7, it must be a rare case, right?
>
> Right, you marked a commit both good and bad. That's probably not very
> common. But it obviously happens. :-)
>
> On 5 January 2018 at 06:28, Yasushi SHOJI <yasushi.shoji@gmail.com> wrote:
>> OK, here is the step to reproduce on git.git
>
> Thank you for providing a script for reproducing this. It helped me come
> up with the attached patch. The patch is based on ma/bisect-leakfix,
> which includes Ævar's patch.
>
> I think this patch could be useful, either as a final "let's test
> something previously non-tested; this would have caught the segfault",
> or simply squashed into Ævar's patch as a "let's add a test that would
> have caught this, and which also tests a previously non-tested code
> path."

Do we really need that?  What is a practical use of a commit having
both good and bad?

Regards,
-- 
            yashi