Le 19/10/2018 à 20:05, Alban Gruin a écrit :
> Le 19/10/2018 à 14:46, SZEDER Gábor a écrit :
>> On Fri, Oct 19, 2018 at 03:02:22PM +0900, Junio C Hamano wrote:
>>> Two large set of topics on "rebase in C" and "rebase -i in C" are
>>> now in 'next'.
>>
>> I see occasional failures in 't5520-pull.sh':
>>
>> […]
>>
>> When running t5520 in a loop, it tends to fail between 10-40
>> iterations, even when the machine is not under heavy load.
>>
>> It appears that these failures started with commit 5541bd5b8f (rebase:
>> default to using the builtin rebase, 2018-08-08), i.e. tip of
>> 'pk/rebase-in-c-6-final', but it's a "flip the big switch" commit, so
>> not very useful...
>>
The bug can be bisected by prepending the command chain of 5520.25 with
`git config --bool rebase.usebuiltin true &&`. Commit 7debdaa4ad
(“builtin rebase: support `--autostash` option”) is the culprit, but
some commits that cause this test to fail with a different error (ie.
“TODO”…) must be marked as good.
The error comes from the call to `git stash apply $stash_id' in
builtin/rebase.c:261. When $stash_id only contains decimals and no
letters, git-stash tries to apply stash@{$stash_id}[0][1]. Thas was not
a real problem with the shell script, because it did not abbreviate the
object id of the stashed commit, so it was very unlikely that the oid
would contain only digits. builtin/rebase.c shortens the oid[2], making
this problem more likely to occur.
We could add a switch to git-stash to tell it that the parameter is an
oid, not anything else.
[0] https://github.com/git/git/blob/master/git-stash.sh#L514-L520
[1] https://github.com/git/git/blob/pu/builtin/stash.c#L167-L168
[2] https://github.com/git/git/blob/next/builtin/rebase.c#L1373
Cheers,
Alban
