rebase --abort Unespected behavior

rebase --abort Unespected behavior

[Blaise Garant]

Hello,

I don't know if this is a bug but it was unexpected for us. I
accidentally added untracked files through a `git add .` while doing
an interactive rebase and aborting the rebase deleted those files. Is
this to be expected?

To reproduce:
mkdir test_folder
cd test_folder
git init
touch first
git add .
git commit -m 'First'
echo 1 >> first
git add .
git commit -m 'Second'
echo 2 >> first
git add .
touch second
git commit -m 'Third'
git rebase -i HEAD~2 #set second to be edited
git add .
git status        #second should have staged
git rebase --abort
ls        #second has been deleted

Not sure this is an expected behavior.
Thanks
Blaise Garant

Re: rebase --abort Unespected behavior

[Phillip Wood]

Hi Blaise

On 28/02/2020 17:36, Blaise Garant wrote:
> Hello,
> 
> I don't know if this is a bug but it was unexpected for us. I
> accidentally added untracked files through a `git add .` while doing
> an interactive rebase and aborting the rebase deleted those files. Is
> this to be expected?

I agree that this is surprising and undesirable but it's not unexpected 
given the way --abort is implemented. 'rebase --abort' calls 'reset 
--hard <branch we're rebasing>' so it will discard all the uncommitted 
changes in the worktree and reset the worktree and index to the branch tip.

The tricky thing with your situation is that the files are tracked at 
the point we call 'reset --hard' as they've been added to the index so 
git feels free to discard them. Perhaps rather than calling 'reset 
--hard' it would be better to use a custom callback with unpack_trees() 
that errors out if there are any paths in the index that are not in 
HEAD, the commit we just picked or the branch tip we're resetting to. If 
we do that we should consider using the same thing for 
'cherry-pick/merge/reset --abort' as well. --autostash potentially 
complicates things as the file might be in the stash but not in the 
other commits but lets not worry about that at the moment.

If your untracked files were ignored then I think 'git add .' would have 
complained or just not added them, but 'git checkout' and 'git merge' 
will happily overwrite ignored files so ignoring them is not always an 
ideal solution.

Best Wishes

Phillip


> To reproduce:
> mkdir test_folder
> cd test_folder
> git init
> touch first
> git add .
> git commit -m 'First'
> echo 1 >> first
> git add .
> git commit -m 'Second'
> echo 2 >> first
> git add .
> touch second
> git commit -m 'Third'
> git rebase -i HEAD~2 #set second to be edited
> git add .
> git status        #second should have staged
> git rebase --abort
> ls        #second has been deleted
> 
> Not sure this is an expected behavior.
> Thanks
> Blaise Garant
> 

Re: rebase --abort Unespected behavior

[Elijah Newren]

Hi Phillip and Blaise,

On Sat, Feb 29, 2020 at 6:30 AM Phillip Wood <phillip.wood123@gmail.com> wrote:
>
> Hi Blaise
>
> On 28/02/2020 17:36, Blaise Garant wrote:
> > Hello,
> >
> > I don't know if this is a bug but it was unexpected for us. I
> > accidentally added untracked files through a `git add .` while doing
> > an interactive rebase and aborting the rebase deleted those files. Is
> > this to be expected?
>
> I agree that this is surprising and undesirable but it's not unexpected
> given the way --abort is implemented. 'rebase --abort' calls 'reset
> --hard <branch we're rebasing>' so it will discard all the uncommitted
> changes in the worktree and reset the worktree and index to the branch tip.

And it's worth noting that if they had done something similar outside
of rebase/merge/cherry-pick/etc.:

git add $UNTRACKED
git reset --hard

then the $UNTRACKED file would be deleted, so this isn't new or
unusual but matches git behavior elsewhere.

> The tricky thing with your situation is that the files are tracked at
> the point we call 'reset --hard' as they've been added to the index so
> git feels free to discard them. Perhaps rather than calling 'reset
> --hard' it would be better to use a custom callback with unpack_trees()
> that errors out if there are any paths in the index that are not in
> HEAD, the commit we just picked or the branch tip we're resetting to. If

Should such a special callback also be used for reset --hard?

Also, this special callback, as stated here, wouldn't work: paths can
exist in a merge that didn't exist in any of the three commits being
threeway merged.  All of the following situations are cases where that
can happen (and there may be more that I'm just not thinking of off
the top of my head):

1) merge in a not-fully clean state.  rebase may disallow this, at
least right now, but merge traditionally hasn't.  I'm not sure
cherry-pick --no-commit disallows this.
2) directory/file or submodule/file or submodule/subdirectory or
regular-file/symlink conflicts. the merge machinery should be free to
rename paths to something that didn't exist on either side so that the
paths from both side can coexist.
3) directory renames.  If one side renamed z/->y/, and the other side
added a new file z/new, the merge should be allowed to move that file
to y/new (depending on the setting of merge.directoryRenames...).
Note that y/new didn't exist on either side of history nor in the
merge base.

> we do that we should consider using the same thing for
> 'cherry-pick/merge/reset --abort' as well. --autostash potentially
> complicates things as the file might be in the stash but not in the
> other commits but lets not worry about that at the moment.

reset --abort?  Not sure what you're referring to here.

Re: rebase --abort Unespected behavior

[Phillip Wood]

Hi Elijah

On 29/02/2020 15:51, Elijah Newren wrote:
> Hi Phillip and Blaise,
> 
> On Sat, Feb 29, 2020 at 6:30 AM Phillip Wood <phillip.wood123@gmail.com> wrote:
>>
>> Hi Blaise
>>
>> On 28/02/2020 17:36, Blaise Garant wrote:
>>> Hello,
>>>
>>> I don't know if this is a bug but it was unexpected for us. I
>>> accidentally added untracked files through a `git add .` while doing
>>> an interactive rebase and aborting the rebase deleted those files. Is
>>> this to be expected?
>>
>> I agree that this is surprising and undesirable but it's not unexpected
>> given the way --abort is implemented. 'rebase --abort' calls 'reset
>> --hard <branch we're rebasing>' so it will discard all the uncommitted
>> changes in the worktree and reset the worktree and index to the branch tip.
> 
> And it's worth noting that if they had done something similar outside
> of rebase/merge/cherry-pick/etc.:
> 
> git add $UNTRACKED
> git reset --hard
> 
> then the $UNTRACKED file would be deleted, so this isn't new or
> unusual but matches git behavior elsewhere.

It's not new but it can be confusing. rm warns if you remove a newly 
added file but reset --hard will happily blow it away

$ touch untracked
$ git add untracked
$ git rm untracked
error: the following file has changes staged in the index:
     untracked
(use --cached to keep the file, or -f to force removal)
$ git reset --hard
HEAD is now at <oid> <subject>

>> The tricky thing with your situation is that the files are tracked at
>> the point we call 'reset --hard' as they've been added to the index so
>> git feels free to discard them. Perhaps rather than calling 'reset
>> --hard' it would be better to use a custom callback with unpack_trees()
>> that errors out if there are any paths in the index that are not in
>> HEAD, the commit we just picked or the branch tip we're resetting to. If
> 
> Should such a special callback also be used for reset --hard?
 >
> Also, this special callback, as stated here, wouldn't work: paths can
> exist in a merge that didn't exist in any of the three commits being
> threeway merged.  All of the following situations are cases where that
> can happen (and there may be more that I'm just not thinking of off
> the top of my head):
> 
> 1) merge in a not-fully clean state.  rebase may disallow this, at
> least right now, but merge traditionally hasn't.  I'm not sure
> cherry-pick --no-commit disallows this.

I'm pretty sure cherry-pick --no-commit will operate on a dirty index, 
I'd forgotten about that

> 2) directory/file or submodule/file or submodule/subdirectory or
> regular-file/symlink conflicts. the merge machinery should be free to
> rename paths to something that didn't exist on either side so that the
> paths from both side can coexist.
> 3) directory renames.  If one side renamed z/->y/, and the other side
> added a new file z/new, the merge should be allowed to move that file
> to y/new (depending on the setting of merge.directoryRenames...).
> Note that y/new didn't exist on either side of history nor in the
> merge base.

Thanks for pointing that out. I'd assumed we'd only check index entries 
in stage 0 so if the user aborts while these entries are unmerged then 
it wouldn't be a problem but if they've partially resolved the merge 
then --abort could fail.

>> we do that we should consider using the same thing for
>> 'cherry-pick/merge/reset --abort' as well. --autostash potentially
>> complicates things as the file might be in the stash but not in the
>> other commits but lets not worry about that at the moment.
> 
> reset --abort?  Not sure what you're referring to here.

I meant revert

I'm not sure what a good way forward is, blindly wiping out newly added 
files is not great from the users point of view but avoiding false 
positives is tricky

Best Wishes

Phillip