Small rerere in rebase regression

Even though I operate with jc/rerere-multi for some time now, I noticed this regression only recently. Watch out for the duplicated "Recorded preimage" message:

gittest@side:1078> strace -f -o /tmp/git-rebase-conflict.strace git rebase -i master error: could not apply 59f92e0... side

When you have resolved this problem, run "git rebase --continue". If you prefer to skip this patch, run "git rebase --skip" instead. To check out the original branch and stop rebasing, run "git rebase --abort".

Recorded preimage for 'a' Recorded preimage for 'a' Could not apply 59f92e09c890c08442eb1df91eaecec3d1656c86... side

According to the strace, the first "Recorded preimage" is from git cherry-pick, the second is from "git rerere" in this section in git-rebase--interactive.sh:

die_with_patch () {
	echo "$1" > "$state_dir"/stopped-sha
	make_patch "$1"
	git rerere
	die "$2"
}

The buglet bisects to

629716d256a792179325c2cc7945bb2d81dda8c2 is the first bad commit
commit 629716d256a792179325c2cc7945bb2d81dda8c2
Author: Junio C Hamano <gitster@pobox.com>
Date:   Thu Jul 30 15:49:18 2015 -0700

    rerere: do use multiple variants
    
    This enables the multiple-variant support for real.  Multiple
    conflicts of the same shape can have differences in contexts where
    they appear, interfering the replaying of recorded resolution of one
    conflict to another, and in such a case, their resolutions are
    recorded as different variants under the same conflict ID.
    
    We still need to adjust garbage collection codepaths for this
    change, but the basic "replay" functionality is functional with
    this change.
    
    Signed-off-by: Junio C Hamano <gitster@pobox.com>

I'm not sure whether the new behavior is a defect in rerere.c or a consequence of the extra rerere call in interactive rebase...

-- Hannes

Johannes Sixt <j6t@kdbg.org> writes:

Interesting. When running an unnecessary "git rerere" (because it already was run and recorded the preimage), we used to be silent. With the updated code, we say something.

At least it seems that we do not record the same preimage as a different variant, so the regression is hopefully merely cosmetic.

I do not think it is a crime to say "git rerere" repeatedly without changing anything, so I do not think "rebase -i"'s call to rerere in die_with_patch is wrong, even though some calls to it seem to be made after seeing a "git merge" fail (which means we know 'rerere' has been run already).

I think all plumbing that can turn an index into unmerged state has a call to git_rerere() at the end (the only obvious exception being "update-index" that lets you stuff higher stage entries to the index), so as long as "rebase -i" uses them correctly, I suspect that die_with_patch shouldn't have to have a call to "git rerere".

Am 23.05.2016 um 21:43 schrieb Junio C Hamano:

I also come to the conclusion that die_with_patch shouldn't have to have a call to "git rerere". die_with_patch can be called after "git cherry-pick", "git merge", "git commit", all of which have their own rerere() invocation.

However, calling "git rerere" after a failed "git commit" may be destructive: it would record a resolution even though the commit has not be completed. Think of an squash commit being aborted because the user notices an error in the last minute. If that error is in a conflict resolution, that wrong resolution would be recorded.

-- Hannes

Johannes Sixt <j6t@kdbg.org> writes:

So, the behaviour change you observed uncovered a small bug in "rebase -i" that was covered by the old limitation of "rerere" that refrained from creating preimage when there already is one?

I think removing the call to "git rerere" there is a safe and sensible thing regardless, but perhaps authors of "rebase -i" had their own reasons. I dunno (it is unlikely I'll have a chance to do blame and digging today).

Hi,

On Mon, 23 May 2016, Junio C Hamano wrote:

I introduced this in ecfe72ff658124ea301ec7b2bb8b987689303685, as a (late) answer to f131dd492f098f9f565df93df13e35c734284590.

Hannes, could you quickly test whether https://github.com/dscho/git/tree/interactive-rebase calls rerere twice, too? (Please call interactive rebase with the GIT_USE_REBASE_HELPER=true to avoid running the interactive rebase twice.)

I have a hunch that it does not call rerere twice, which would be a nice bonus in that patch thicket (so far, I split the patches into seven patch series, in addition to the ones I already sent, still have to do one thorough review before I'll continue submitting them).

Ciao, Dscho

Am 24.05.2016 um 15:18 schrieb Johannes Schindelin:

It prints the message only once:

gittest@master:1007> GIT_USE_REBASE_HELPER=true git rebase -i side error: could not apply fa62fea... master

When you have resolved this problem, run "git rebase --continue". If you prefer to skip this patch, run "git rebase --skip" instead. To check out the original branch and stop rebasing, run "git rebase --abort".

Recorded preimage for 'a' Could not apply fa62fea... mastergittest@master|REBASE-i 1/1:1008> 5~

(Take note of the missing LF at the end of the message.)

Can this result be interpreted as another indication that the "git rerere" call in die_with_patch can be removed, or are the two git-rebase implementations too different to be comparable?

-- Hannes

Hi Hannes,

On Tue, 24 May 2016, Johannes Sixt wrote:

Thanks for confirming.

Oops. Good catch, thank you! I just fixed this:

	https://github.com/dscho/git/commit/0393d7bb2d

The code used by the rebase--helper is not *quite* the same as the entire cherry-pick code path. For one, we run the sequencer directly, without any of cherry-pick's option parsing.

Having said that, yes, we inherit sequencer's rerere() call that was part of cherry-pick originally ever since aa1a011 (Make cherry-pick use rerere for conflict resolution., 2008-08-10) (amd which was moved to builtin/revert.c in 81b50f3 (Move 'builtin-*' into a 'builtin/' subdirectory, 2010-02-22) and refactored out into the sequencer in 043a449 (sequencer: factor code out of revert builtin, 2012-01-11)).

I did notice that rerere() call when implementing rebase -i's functionality in the sequencer, and dropped the extra call in error_with_patch() (the libified version of rebase -i's die_with_patch function).

In short: yes, the explicit `git rerere` call can be dropped, that is essentially what I did in the rebase--helper branch.

Ciao, Dscho

Hi Hannes,

On Wed, 25 May 2016, Johannes Schindelin wrote:

Argh. I fixed it in https://github.com/dscho/git/commit/0704d6f425 of course.

Ciao, Dscho

Am 25.05.2016 um 07:38 schrieb Johannes Schindelin:

Here's a patch to do that.

--- 8< ---
Subject: [PATCH] rebase -i: remove an unnecessary 'rerere' invocation

Interactive rebase uses 'git cherry-pick' and 'git merge' to replay commits. Both invoke the 'rerere' machinery when they fail due to merge conflicts. Note that all code paths with these two commands also invoke the shell function die_with_patch when the commands fail.

Since commit 629716d2 ("rerere: do use multiple variants") the second operation of the rerere machinery can be observed by a duplicated message "Recorded preimage for 'file'". This second operation records the same preimage as the first one and, hence, only wastes cycles. Remove the 'git rerere' invocation from die_with_patch.

Shell function die_with_patch can be called after the failure of "git commit", too, which also calls into the rerere machinery, but it does so only after a successful commit to record the resolution. Therefore, it is wrong to call 'git rerere' from die_with_patch after "git commit" fails.

Signed-off-by: Johannes Sixt <j6t@kdbg.org>
---
 git-rebase--interactive.sh | 1 -
 1 file changed, 1 deletion(-)

Hi Hannes,

On Fri, 27 May 2016, Johannes Sixt wrote:

ACK

Ciao, Dscho

Am 24.05.2016 um 00:11 schrieb Junio C Hamano:

I had a closer look, and found that the failure mode I described does not occur: after having resolved a commit, the user can call "git rebase --continue", but the "git commit" invocation that occurs there is guarded with a mere 'die', not 'die_with_patch'.

All other "git commit" guarded with 'die_with_patch' are either not interactive or are after a non-conflicting merge or cherry-pick (and so there was no opportunity to resolve a conflict).

-- Hannes