From: Mikhail Filippov <mikhail@filippov.me>
Date: Mon, 12 Sep 2016 08:38:00 GMT
Subject: Re: [PATCH] Do not record unstaged deleted file upon recursive merge if file was moved outside of working tree with enabled sparse-checkout.
Message-ID: <798801473669480@web11o.yandex.ru>
In-Reply-To: <xmqqeg4q6mw4.fsf@gitster.mtv.corp.google.com>

It's a very unexpected behaviour when a user sees a deleted file after a merge with enabled sparse-checkout. Moreover, when the user resolves merge conflicts and commits the changes with the command "git commit -am xxx", a repository can be broken because all the moved files will be deleted. Finally, it's really hard to find a user who deleted these files because "git log file" doesn't show any merge commits by default. I'm not sure that my fix is correct but I checked all tests and I didn't find a better way to prevent files deleting.

12.09.2016, 02:24, "Junio C Hamano" <gitster@pobox.com>:
>    Mikhail Filippov <mikhail@filippov.me> writes:
>
>>     ---
>
>    You'd need a lot more explanation on why this is needed
>    (i.e. without it what behaviour you would get, and why that
>    behaviour is wrong).
>
>>      merge-recursive.c | 9 +++++---
>>      t/t6042-merge-rename-corner-cases.sh | 42 ++++++++++++++++++++++++++++++++++++
>>      2 files changed, 48 insertions(+), 3 deletions(-)
>>
>>     diff --git a/merge-recursive.c b/merge-recursive.c
>>     index e349126..25dc701 100644
>>     --- a/merge-recursive.c
>>     +++ b/merge-recursive.c
>>     @@ -1724,9 +1724,12 @@ static int merge_content(struct merge_options *o,
>>                       */
>>                      path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
>>                      if (!path_renamed_outside_HEAD) {
>>     - add_cacheinfo(o, mfi.mode, &mfi.oid, path,
>>     - 0, (!o->call_depth), 0);
>>     - return mfi.clean;
>>     + struct stat st;
>>     + if (lstat(path, &st) == 0) {
>>     + add_cacheinfo(o, mfi.mode, &mfi.oid, path,
>>     + 0, (!o->call_depth), 0);
>>     + return mfi.clean;
>>     + }
>
>    I cannot guess what you are trying to achieve without explanation in
>    the proposed log message, but I can say that this unconditional
>    checking of a working tree file cannot be correct (there may or may
>    not be other things that are wrong with this change, which cannot be
>    judged without more information).
>
>    Imagine your o->call_depth is not zero, i.e. we are making a virtual
>    common ancestor with this merge, in which case any of the three
>    trees involved may have nothing to do with the current working tree
>    files.
>
>>     +test_expect_success 'move file/sparse-checkout/merge should not delete moved file' '
>>     + git rm -rf . &&
>>     + git clean -fdqx &&
>>     + rm -rf .git &&
>>     + git init &&
>
>    Yuck. This is inherited from existing tests but I think they need
>    to be cleaned up. It is not your fault, and it is not in the scope
>    of this change.
>
>>     + git status >output &&
>>     + cp output /tmp/a &&
>
>    Huh?
>
>>     + test_i18ngrep "nothing to commit" output
>>     +'
>>     +
>>      test_done




-- 
Mikhail Filippov 
Software Developer
JetBrains
http://jetbrains.com
“The Drive To Develop"